Cooperative Optimization Framework for Video Resource Allocation with High-Dynamic Mobile Terminals

Abstract

Under the typical scenario of high-speed mobility, channel disturbances at the physical layer may disturb the transmission of video base layers. Due to the close dependency of Scalable Video Coding (SVC) on base layers, such disturbances will result in retransmissions and handover delays. Meanwhile, ineffective enhancement layers continue to occupy resources, ultimately causing system performance collapse and further exacerbating physical-layer disturbances. To address this challenge, we propose an edge computing resource coordination optimization scheme for highly dynamic mobile terminals. The scheme first empowers the SVC layered transmission with the local caching capabilities, enabling rapid retransmission of base layer data by employing a Lyapunov optimization framework for transmission queue scheduling. Secondly, we design a strategy for dynamically releasing the enhancement layer (EL) cache. This can mitigate resource waste caused by invalid enhancement layers. Finally, Lyapunov drift optimization is implemented to ensure base layer transmission stability and accelerate system state convergence. Simulation and experimental results demonstrate that the proposed scheme significantly improves video transmission reliability and user experience in highly dynamic network environments.

1. Introduction

With the rapid deployment of 6G networks, satellite Internet, and intelligent edge computing, the demand for real-time applications in highly dynamic mobile communication networks (e.g., vehicular communication systems, high-speed rail networks, and urban subway infrastructures) is surging [1]. In these scenarios, end-users encounter critical challenges including rapid channel state variations, ultra-high mobility speeds, and heterogeneous resource competition, rendering traditional communication networks inadequate for meeting stringent requirements of ultra-low latency, high reliability, and energy efficiency optimization [2]. According to the International Telecommunication Union (ITU), global mobile data traffic is projected to grow at a compound annual rate exceeding 30% by 2030, with real-time video streaming dominating network traffic [3]. In such highly dynamic mobile environments, the cooperative allocation of computing, communication, and storage resources to guarantee end-user Quality of Service (QoS) while optimizing energy consumption has emerged as a key technical challenge [4].

Traditional mobile edge computing (MEC) systems substantially reduce latency and enhance user experience by decentralizing computing resources to the network periphery. However, existing MEC resource allocation strategies predominantly target static or low-dynamic environments that struggle to adapt to the rapid fluctuations in channel states and heterogeneous user demands inherent in highly dynamic mobile communication networks [5]. Furthermore, conventional fixed resource allocation or scheduling approaches relying on simplistic heuristic mechanisms frequently result in suboptimal resource utilization or service disruptions when confronted with the stochastic access patterns and variable resource requirements of highly mobile users [6]. For instance, ref. [7] employed Dinkelbach’s method integrated with convex optimization techniques to develop an online algorithm that maximizes energy efficiency while maintaining queue stability for static or slowly varying wireless channels through the hybrid cooperative mechanism of Backscatter Communication (BackCom); ref. [8] proposed a Genetic Algorithm (GA) and Heuristics (MATS)-based framework for traditional task scheduling and resource allocation to optimize task offloading latency in mobile edge computing environments.

There are few studies on real-time video transmission in highly dynamic mobile communication networks, especially those that address the complex challenges of resource optimization in such environments. The work [9] presents a semantic communication framework based on Dynamic Decision Generation Networks (DDGNs) and Generative Adversarial Networks (GANs), which achieves high-compression, low-distortion key-frame transmission for video streams in hyper-dynamic mobile networks through dynamic feature compression and adversarial reconstruction optimization. Ref. [10] introduces an SDN-based framework for centralized management of VR video resources in 6G cellular systems, ensuring seamless low-latency VR experiences under rapidly changing network conditions by dynamically reallocating bandwidth and computational tasks. Ref. [11] proposes an amplified programmable hypersurface system with joint modulation capabilities to synchronize real-time video transmission with wireless energy transfer in complex electromagnetic environments via dynamic beamforming and joint modulation strategies, thereby addressing stability and energy efficiency challenges in highly dynamic scenarios. Ref. [12] develops an intelligent tracking system combining computer vision and programmable hypersurface technologies, enabling real-time video transmission for moving targets in dynamic environments through real-time target sensing and adaptive beamforming. Despite these contributions, a significant research gap remains in the area of real-time video transmission under high-mobility and hyper-dynamic network conditions. Specifically, most existing schemes predominantly focus on downlink communication layer optimization, while there is limited investigation into cross-layer co-optimization frameworks and uplink-oriented task offloading and resource allocation mechanisms that are specifically designed to meet the stringent latency, reliability, and adaptability requirements of such environments. This lack of holistic, adaptive, and bidirectional optimization strategies highlights the need for further research on intelligent and dynamic resource management in real-time video transmission systems for future high-mobility networks.

Under high-mobility scenarios, significant channel fluctuations induced by Doppler shift may trigger transmission interruptions in the base layer (BL) of video streams. Given the close dependency of Scalable Video Coding on BL integrity, where BL loss renders all enhancement layers (ELs) ineffective [13], localized physical layer disturbances can propagate into systemic transmission challenges: channel fluctuations activate BL retransmission while interacting with base station handover latency, causing dramatic increases in end-to-end latency [14]. This creates a closed-loop deterioration pathway: physical layer disturbances, dependency amplification, resource contention, performance collapse, and intensified physical layer disturbances [15].

To address the challenges in high-dynamic mobile communication networks, we propose a user-centric resource coordination scheme that leverages the SVC layered transmission architecture and edge computing. Upon detecting channel degradation, the proposed scheme retransmits cached BL content, reducing backhaul latency. A Lyapunov optimization framework manages transmission queues, balancing retransmission rates and handover strategies to minimize end-to-end latency. Within Lyapunov optimization, edge nodes allocate dedicated resources for BL and EL using MEC systems [16]. If BL loss invalidates EL data, corresponding resources are reallocated to prioritize BL transmission stability. The MEC system also adapts EL redundancy and compression ratios to minimize bandwidth and computational consumption under poor channel conditions. To address physical layer disturbances, two virtual queues namely delay disturbance and resource occupancy are introduced [17]. The delay queue compensates for channel jitter by dynamically adjusting weights based on BL packet loss and retransmission delays. The resource queue minimizes Lyapunov drift, optimizing the use of edge computing resources [18]. Finally, a Hierarchical Quantum Particle Swarm Optimization (HQPSO)-based algorithm is introduced for joint offloading and resource allocation. This algorithm rapidly identifies near-optimal solutions, ensuring system stability and preventing excessive retransmissions or handovers caused by sudden performance degradation.

The remainder of this paper is structured as follows: Section 2 presents the system model formulation and problem definition; Section 3 elaborates on the joint offloading and resource allocation optimization algorithm design; Section 4 describes the simulation experiment configurations and performance analysis; Section 5 summarizes the key research contributions and proposes future research directions.

2. System Model

As shown in Figure 1, in this paper, we propose a collaborative architecture that integrates SVC, MEC, and Lyapunov for real-time video streaming transmission requirements in highly dynamic network environments. $h_{u,m}^j\left(t\right)$ represents the channel state; $Q_u\left(t\right)$ represents the queue backlog; ${\lambda }_u\left(t\right)$ represents the task arrival rate; $\chi \left(t\right)\in {\{0,1\}}^3$ represents the offloading decision; $f_{u,m}^l\left(t\right)\ge 0$ represents the resource allocation; $p_u\left(t\right)\in [0.1,1.0]w$ represents the power control.

The system architecture comprises three functional components: a mobile client, a network resource layer, and a dynamic control layer [19]. The mobile client implements SVC-based decomposition of video streams into the base layer and enhancement layers with hierarchical dependencies, and establishes a dynamic virtual queue to enable real-time feedback of task backlog status. The base layer is preferentially offloaded to macro base stations through ultra-reliable low-latency communication (URLLC) links, while the enhancement layers dynamically select Orthogonal Frequency Division Multiple Access (OFDMA) sub-bands for transmission to small base stations based on channel state prediction [20]. The network resource layer integrates heterogeneous computing resource pools from macro/small base stations, performs on-demand resource allocation through elastic scaling mechanisms, and enforces layered dependency constraints to ensure video transmission integrity [21]. The dynamic control layer incorporates Lyapunov optimization modules to jointly optimize offloading policies, resource allocation weights, and power control parameters in real time, achieving dynamic balance between user experience quality maximization and queue stability through drift-plus-penalty minimization [22].

Within the SVC–MEC–Lyapunov framework, we instantiate a virtual-queue layer that serves as the system’s control memory, closing the loop from state observation to action. This layer fuses measurements of channel, queue, and cache states with online control of task offloading, sub-band assignment and power allocation, and edge-compute provisioning, and it actuates outcomes such as HARQ retransmissions for the base layer and eviction of stale enhancement layers. At each time slot t, the controller dynamically updates two interdependent virtual queues: (i) a distortion queue, defined per user u and layer l, which accounts for channel jitter and compensates for the dependencies between BL and EL transmissions; (ii) a task queue, maintained per user, which jointly captures the utilization of radio resources and edge computing capacity. The SVC layer hierarchy, together with the heterogeneous MBS/SBS infrastructure, informs the weighting mechanism to prioritize BL reliability and ensure the timely eviction of obsolete or invalid EL data. This integrated design effectively transforms the system into a closed-loop scheduling architecture, as opposed to a best-effort open-loop pipeline, thereby enabling rigorous performance guarantees and dynamic adaptability to time-varying network conditions.

Specifically, the queue updates follow:

$$Z_u^l(t+1)=max\left[Z_u^l\left(t\right)-{\rho }_u^l{\mu }_u^l\left(t\right),0\right]+{ϵ}_u^l\left(t\right).$$

(1)

where $Z_u^l\left(t\right)$ represents the distortion queue state for user u’s video layer l at time slot $t, {\rho }_u^l$ is the priority weight factor, ${\mu }_u^l\left(t\right)$ is the successfully transmitted data amount, and ${ϵ}_u^l\left(t\right)$ accounts for distortion accumulation due to transmission failures.

The task queue is updated as follows:

$$Q_u(t+1)=max\left[Q_u\left(t\right)-\sum _{m,l}{\mu }_{u,m}^l\left(t\right),0\right]+A_u\left(t\right).$$

(2)

where $Q_u\left(t\right)$ denotes the task queue length for user $u,{\mu }_{u,m}^l\left(t\right)$ is the data transmitted via base station m, and $A_u\left(t\right)$ is the new arriving video task volume. This design prioritizes high-urgency tasks in resource contention, particularly for latency-critical applications.

$U=\{1,2,3,\dots ,U\}$ is used to denote the users in the mobile video system, and each user generates a real-time video stream with frame rate $F_u$ and resolution $R_u$. $M=\{1,2,3,\dots ,M\}$ denotes the set of MEC servers, where $m=0$ is a macro base station, $m\ge 1$ is a small base station, and the computational power $f_m^{max}$ is heterogeneously distributed with the coverage radius $D_m$. $L=\{1,2,\dots ,L\}$ denotes the set of SVC video tiers. $\mathcal{J}=\{1,2,\dots ,J\}$ denotes the set of orthogonal sub-bands with bandwidth $W=180 \mathrm{kHz}$, supporting OFDMA multiple access.

2.1. User Side Mode

Assume that for each user $u\in U$ at the user terminal, there is one and only one video computation task, denoted as $T_u$, which is atomic and cannot be divided into subtasks. The performance of each computational task $T_u$ is expressed as a tuple consisting of two descriptions $〈d_u,c_u〉$, where $d_u$ denotes the size of the task data transmitted from the user-side device to the MEC server, and $c_u$ denotes the size of the resources required to complete the computational task, both of which can be obtained based on the size of the user’s task execution data volume. In the MEC system of this paper, each computational task can perform video parsing at the user end or can be offloaded to the MEC server to perform video parsing. By offloading the video computation task to the server, the video user saves the energy required for the computation task, but sending the video task to the uplink adds more time and energy [23].

Using $f_u^{\mathrm{local}}$ to denote the local computing power of user $u, f_u^{\mathrm{local} }> 0$, in terms of the number of CPU cycles per second, if the user u performs video parsing locally, the latency to complete the task is as follows:

$$t_u^{\mathrm{local}} = {\displaystyle \frac{c_u}{f_u^{\mathrm{local}}}}.$$

(3)

.

The energy consumption model is used to represent the energy consumed by the user to parse the video locally. Using f for the CPU frequency and $\alpha$ for the energy factor, each computation cycle is $\epsilon =\alpha f$, where the size of $\alpha$ is determined by the chip architecture. According to the above model, the energy consumption for locally executing the video task $T_u$ is as follows:

$$E_u^{\mathrm{local}} = \alpha {\left(f_u^{\mathrm{local}}\right)}^2{c}_u.$$

(4)

The user equipment is based on SVC technology, which structures the real-time captured video stream in time slices. Each time slice lasts for 1 s and corresponds to $F_u$ frames of video data, which are generated into layered packets by the H.265/SVC encoder: BL contains 1 frame with NOTICE P frames, with a bit rate of $B_u^1=K{R}_u{F}_u$, where K is the compression factor, which determines the minimum acceptable quality of the video, and the enhancement layer (EL) is realized by layered incremental coding, where the bit rate of the first layer is the enhancement rate, providing resolution enhancement or dynamic range extension. The EL is realized by layered incremental coding, where the code rate of layer one $B_u^l=B_u^{l-1}(1+\eta )$ is the enhancement rate, which provides resolution enhancement or dynamic range extension. This hierarchical structure allows users to flexibly choose the transmission layer according to the network conditions.

2.2. Task Offloading Model

Assuming a multi-user multi-MEC server architecture where each user’s video computation tasks can be selectively offloaded to any available small base station within the system [5], three distinct latency components emerge during the offloading process: (i) uplink transmission latency when offloading video tasks to MEC servers; (ii) computation processing latency at the base station’s MEC server; and (iii) downlink transmission latency for returning computation results to the user [24]. Given that uplink data size is typically significantly smaller than downlink data, and considering the inherent asymmetry in wireless channel capacity where downlink data rates substantially exceed uplink rates, the downlink transmission delay can be neglected in computational complexity analysis [25].

Similar to the work [26], this paper applies the OFDMA technique to the uplink transmission system by dividing the transmission band B into W equal sub-bands of size N, i.e, $W=B/N\left[\mathrm{Hz}\right]$, and each BS (Band Width) can receive up to one user’s upload task at the same time and can receive upload tasks from N users simultaneously. Assuming that the set of available sub-bands for each BS is $N=\{1,2,3,\dots ,N\}$, the offloading variable is defined as $x_{um}^{jl}$ considering the allocation of uplink sub-bands, where $u\in U,m\in M,j\in N,l\in L$; $x_{um}^{jl}=1$ denotes that task $T_u$ offloads the l layer of the video from user u to base station m via sub-band j, and $x_{um}^{jl}=0$ has the opposite meaning, i.e., the following:

$$x_{um}^{jl}=\left\{\begin{array}{cc}1,\hfill & \mathrm{User} \mathrm{offloads} \mathrm{video} \mathrm{to} \mathrm{base} \mathrm{station} \mathrm{via} \mathrm{subband}\hfill \\ 0,\hfill & \mathrm{else}\hfill \end{array}\right.$$

(5)

Assuming that the task offloading policy is X, then $X=\left\{x_{um}^{jl}\mid x_{um}^{jl}=1,\forall u\in U,\right.$$\left.m\in M,j\in N,l\in L\right\}$. In the system of this paper, each video task can either be parsed locally or offloaded to an associated MEC for parsing. Therefore, the feasibility analysis leads to the following:

$$\sum _{m\in M}\sum _{j\in N}{x}_{um}^{jl}\le 1,\forall u\in U,l\in L.$$

(6)

Furthermore, assuming that both the user side and the BS are equipped with a single antenna for uplink transmission, and the power of user u to transmit a task to the BS is $P_u\left[W\right]$, then $P=\left\{P_u\mid 0<P_u\le P_u,u\in U\right\}$, denoting the power pooling. Due to the application of OFDMA technology in the uplink, users of the same base station will transmit tasks on different sub-bands, which suppresses the mutual interference among sub-bands [27]. However, there is still interference between the mobile devices, where the Signal–Noise Ratio (SINR) of the user u uploading the task to the sub-band j is calculated as follows:

$${\gamma }_{um}^{jl}={\displaystyle \frac{p_u{h}_{um}^j}{{\sum }_{r\in M\setminus \left\{m\right\}}{\sum }_{k\in U_r}{x}_{kr}^{jl}{p}_k{h}_{km}^j+{\sigma }^2}},\forall u\in U,m\in M,j\in N,l\in L$$

(7)

In the formula, ${\sigma }^2$ denotes the background noise variance. $h_{um}^j$ represents the channel gain coefficient between the base station (BS) and associated users for transmission. $p_u$ indicates the transmit power of user u when offloading tasks to the server. $x_{kr}^{jl}$ signifies user k uploading the $l_{-}th$ layer of a video through subcarrier j to server m. Furthermore, $p_k$ stands for the transmit power of user k in the process of offloading tasks to the server, and $h_{km}^j$ refers to the channel gain coefficient between server $BSM$ and user k for transmission.

The path loss model adopted in this paper [28] is given by $140.7+36.7·lg{d}_{um}^j$, where $d_{um}^j$ represents the distance between BSM and user u (in units of km). Each user’s video task is transmitted on only one subcarrier; therefore, the rate at which user u uploads video to server BSM [bits/s] is expressed as follows:

$$R_{um}\left(\chi \right)=W{log}_2\left(1+{\gamma }_{um}\right).$$

(8)

In the formula, ${\gamma }_{um}={\sum }_{j\in N}{\gamma }_{um}^{jl}$, where ${\gamma }_{um}^{jl}$ denotes the signal-to-noise ratio (SNR) from user u to server BSM on subcarrier j. Consequently, the transmission time for user u to send video task $d_u$ over the uplink is given by the following equation:

$$t_{up}^u=\sum _{m\in M}{\displaystyle \frac{x_{um}^{jl}{d}_u}{R_{um}\left(\chi \right)}},\forall u\in U$$

(9)

In the equation, $x_{um}={\sum }_{j\in N}{\sum }_{l\in L}{x}_{um}^{jl}$, where $x_{um}^{jl}$ represents the offloading of the l-th layer of a video from user u to server m via subcarrier j.

While task offloading provides significant benefits, it is equally important to address how to efficiently allocate computational resources at the MEC layer. In the following subsection, we will explore the SVC-MEC computing resource integration model, which optimizes video transmission and processing at the edge.

2.3. SVC-MEC Computing Resource Integration Model

In dense heterogeneous network environments, this paper formulates a dynamic MEC resource scheduling model tailored for multi-user real-time video streaming demands by leveraging SVC hierarchical characteristics. The proposed model achieves efficient computing resource allocation and QoS guarantees through synergistic integration of multi-BS resource constraints and SVC hierarchical features. Within the system architecture, MBSs and SBSs are provisioned with differentiated computing resource pools [29], where MBSs prioritize SVC BL tasks by reserving $20\%$ of the resource pool $f_0^{\mathrm{reserved} }= 0.2$ $f_s^0$ and implementing a lightweight containerized instance preloading mechanism. The cold-start latency of the base layer tasks is compressed to 5 ms to ensure the real-time requirement $t_{\mathrm{exec} }^1\le 100 \mathrm{ms}$, while the small base station focuses on the resilient processing of the EL tasks by adopting a dynamic resource allocation mechanism based on the SVC hierarchical dependency:

$$f_{u,m}^l\left(t\right)={\displaystyle \frac{y_u^l\left(t\right)·B_u^l}{{\sum }_k{y}_k^l\left(t\right)B_k^l}}·f_s^m,y_u^l\left(t\right)\in \{0,1\}.$$

(10)

In the equation, $f_{u,m}^l\left(t\right)$ denotes the computational resources allocated by base station m to user u for the $l_{-}$th layer of video at time slot t. The numerator, $B_u^l$, represents the bit rate of the $l_{-}$th layer of the video. The denominator is the total bit rate of all users’ tasks at the same layer. Additionally, $f_s^m$ signifies the total computational resource capacity of base station m.

Activate high-level resource allocation only when the completion of a low-level task is detected, and introduce a dynamic fallback mechanism as shown below to prevent resource overload:

$$\sum f_{u,m}^l>f_s^m\Rightarrow \mathrm{Suspension} \mathrm{of} \mathrm{the} \mathrm{top} \mathrm{EL} \mathrm{mandate} .$$

(11)

In the equation, $\sum f_{u,m}^l$ represents the total amount of computational resources already allocated by base station m. Here, $f_s^m$ denotes the total computational resource capacity of base station m. The highest EL task refers to the video stream task with the highest level in the enhancement layer.

For bursty traffic scenarios, the model is designed with an elastic resource expansion mechanism:

$$f_s^m\left(t\right)=f_s^m\left(1+\beta tanh\left[{\displaystyle \frac{\sum Q_u\left(t\right)}{U{Q}_{\mathrm{avg}}}}\right]\right),$$

(12)

where $\beta$ is the elasticity expansion coefficient, $tanh(·)$ is the hyperbolic tangent function, which is used to smooth the adjustment of the resource expansion amplitude, and $Q_{\mathrm{avg}}$ represents the average queue length hole value of the system.

The capacity of the small base station resource pool is dynamically adjusted to cope with the instantaneous load surge, and, at the same time, a rapid response mechanism is established to automatically trigger the hierarchical degradation strategy when resource overload is detected to ensure system stability.

The integration of SVC and MEC resources establishes a solid foundation for efficient video transmission and computation at the edge. However, to maintain system stability and guarantee performance under dynamic network fluctuations, a more rigorous optimization framework is required. Therefore, the following subsection introduces the Lyapunov optimization model, which provides a systematic approach to dynamically balance queue stability, delay, and resource utilization.

2.4. Lyapunov Optimization Model

In highly dynamic network environments, resource allocation for real-time video streaming confronts multiple challenges including rapidly fluctuating channel conditions and drastic variations in user demand [30]. Conventional static optimization approaches struggle to adapt to these time-varying characteristics, while prediction-driven algorithms face limitations in computational complexity and forecasting accuracy. The Lyapunov optimization framework offers a comprehensive theoretical foundation for addressing this challenge—it characterizes system dynamics through virtual queue construction and converts complex long-term stochastic optimization problems into deterministic subproblems using the drift-plus-penalty methodology [6]. The following analysis systematically explores the engineering implementation of this framework across dynamic adjustment mechanism for drift-plus-penalty optimization.

The choice of parameter V is one of the core challenges of the Lyapunov framework. Traditional static settings (e.g., fixed) are difficult to adapt to network load breaking. For this reason, the adaptive V regulation algorithm is proposed, based on the Lyapunov function:

$$L\left(\theta \left(t\right)\right)={\displaystyle \frac{1}{2}}\sum _u\left[Q_u^2\left(t\right)+\sum _{l=1}^L{\left(Z_u^l\left(t\right)\right)}^2\right].$$

(13)

In the equation, $L\left(\theta \right(t\left)\right)$ represents the Lyapunov function value at time slot t; $Q_u^2\left(t\right)$ quantifies the backlog level of the task queue; ${\left(Z_u^l\left(t\right)\right)}^2$ signifies the cumulative effect of distortion in video layers. This function amplifies the penalty weight for large queue states through a quadratic term, encouraging the system to prioritize high-backlog tasks. A smaller value indicates superior system stability. The conditional drift is expressed by the following equation:

$$\mathsf{\Delta }\left(\theta \right(t\left)\right)=E\left[L\right(\theta (t+1))-L(\theta \left(t\right))\mid \theta (t\left)\right]$$

(14)

Specific adjustment strategies include the following: (i) Short-term adjustment: dynamically scale V based on the ratio of instantaneous queue length to distortion value. For example, when ${max}_u{Q}_u\left(t\right)/Q_{\mathrm{avg}} > 2$, temporarily reduce V to prioritize stabilizing the queue. (ii) Long-term learning: utilize reinforcement learning (such as DQN) to train the adjustment strategy for V, with a reward function based on long-term Quality of Experience (QoE) and delay metrics.

This dynamic regulation enables the system to automatically switch to low-latency mode during congested periods (such as live sports broadcasts), while enhancing video quality during network idle times (such as late at night). Experiments show that the adaptive V strategy can improve QoE by 15∼20% compared to fixed-value schemes.

2.5. Systematic General Computational Model

Based on the description of the modules above, it is known that in a video processing system, each user device generates different video computing tasks, which usually have different computing resource requirements $R_i$ and data transmission requirements $D_i$. These tasks may be processed locally or offloaded to the MEC (mobile edge computing) servers for computation over the wireless network. The system needs to make a decision on whether to offload a task based on the computing power of the device $C_{\mathrm{local}}$ and the network condition. To this end, the system takes into account multiple factors, including computing power, transmission delay, and network bandwidth, and makes a dynamic judgment. The core goal of offloading decision-making is to improve the overall efficiency of the system by minimizing delay and energy consumption, while ensuring the resource load balance of the system [31]. In this context, the offloading decision is calculated by the following formula:

$$\mathrm{Decision}\left(T_i\right)=\left\{\begin{array}{cc}\mathrm{MEC},\hfill & \mathrm{if} C_{\mathrm{MEC}}>R_i,T_i\le \mathrm{Threshold}\hfill \\ \mathrm{Local},\hfill & \mathrm{if} C_{\mathrm{local}}>R_i,T_i>\mathrm{Threshold}\hfill \end{array}\right.$$

(15)

In the formula, $C_{MEC}$ and $C_{\mathrm{local}}$ represent the computational capabilities of the MEC server and user devices, respectively. $\left(T_i\right)$ denotes the data transmission delay for task $T_i$, with a threshold used to determine whether offloading the task to the MEC server would enhance performance. This decision-making process aids in determining the optimal processing method for tasks [27], ensuring that computational tasks are completed within a reasonable timeframe while avoiding system overload due to insufficient network transmission or local computing resources.

To achieve dynamic scheduling and optimization of tasks under highly dynamic scenarios, the system employs a Lyapunov optimization framework to manage resource allocation. This framework adjusts in real-time based on changes in the task queue $Q_{i,l}\left(t\right)$, which represents the queue state of the $l_{-}th$ layer for the $i_{-}th$ type of task at time t. The system’s objective is to adjust resource allocation according to the arrival and processing status of each task, minimizing system latency and energy consumption while ensuring balanced system load [32]. The queue evolution within the Lyapunov optimization framework is described by the following equation:

$$Q_{i,l}(t+1)=Q_{i,l}\left(t\right)+\left(A_{i,l}\left(t\right)-D_{i,l}\left(t\right)\right).$$

(16)

In the formula, $A_{i,l}\left(t\right)$ represents the arrival rate of tasks at time t, and $D_{i,l}\left(t\right)$ denotes the processing rate of the task queue at time t. The dynamic adjustment of queue states ensures that the system can optimize resource allocation based on the current task load.

During the optimization process of resource allocation, the system aims to minimize the drift-plus-penalty function, ensuring that tasks are processed according to their priority order while avoiding excessive delays and resource wastage. This objective function is expressed by the following equation:

$$V\left(t\right)=\sum _{i,l}\left(Q_{i,l}\left(t\right)·{\rho }_{i,l}\right)+\sum _{i,l}\left(A_{i,l}\left(t\right)·{\lambda }_{i,l}\right),$$

(17)

where ${\rho }_{i,l}$ is the weight of the task, and ${\lambda }_{i,l}$ is the drift penalty coefficient. By regulating these values, the system is able to efficiently allocate computational resources, avoiding a certain portion of resources being over-occupied and ensuring the optimization of overall performance.

The primary objective of this research is to synergistically optimize offloading decisions and resource allocation for video computing tasks in hyper-dynamic environments. Video stream processing requires ensuring both data integrity and quality while minimizing transmission and computational latency [33]. Latency optimization constitutes a critical system design dimension, particularly for real-time video streaming applications where the system must guarantee rapid response capabilities and timely task execution [34]. To achieve this, the system dynamically adjusts computational resource allocation through real-time monitoring of BL and EL task latencies, thereby minimizing overall task completion time [35]. The mathematical formulations for BL latency and EL latency are specified as Equations (18) and (19):

$${\mathrm{Delay}}_{BL}\left(t\right)={\displaystyle \frac{R_{BL}\left(t\right)}{B_{\mathrm{local}}}} + {\displaystyle \frac{R_{BL}\left(t\right)}{B_{MEC}}}$$

(18)

$${\mathrm{Delay}}_{EL}\left(t\right)={\displaystyle \frac{R_{EL}\left(t\right)}{B_{\mathrm{local}}}} + {\displaystyle \frac{R_{EL}\left(t\right)}{B_{MEC}}}$$

(19)

In the public center, $B_{\mathrm{local}}$ and $B_{MEC}$ denote the bandwidth of the local device and the MEC server, respectively, while $R_{BL}\left(t\right)$ and $R_{EL}\left(t\right)$ are the transmission demands of the base layer and the enhancement layer. The system dynamically adjusts the bandwidth allocation and optimizes the transmission path according to these demands, thus reducing the overall delay and improving the efficiency of video stream processing.

While ensuring a real-time response, the system must also minimize energy consumption. This not only helps extend the life of the equipment but also improves the overall stability of the system. During task processing, the system dynamically adjusts the energy allocation according to the use of different computing resources to ensure a balance between energy consumption and latency. The energy consumption E can be calculated by the following formula:

$$E\left(t\right)=P_{\mathrm{local}}·{\mathrm{Delay}}_{\mathrm{local}}\left(t\right)+P_{\mathrm{MEC}}·{\mathrm{Delay}}_{\mathrm{MEC}}\left(t\right)$$

(20)

In the formula, $P_{\mathrm{local}}$ and $P_{\mathrm{MEC}}$ represent the energy consumption of local devices and MEC servers, respectively. ${Delay}_{\mathrm{local}}\left(t\right)$ and ${Delay}_{\mathrm{MEC}}\left(t\right)$ denote the latency for local and MEC processing. By optimizing latency and energy consumption, the system can achieve more efficient resource management.

Based on the above multi-dimensional modeling, the system optimization objective is defined as maximizing the user’s comprehensive QoE under the premise of ensuring queue stability, and the system implements a dynamic resource management and scheduling framework. This framework continuously adjusts task offloading, resource allocation, load balancing, coding optimization, and delay and energy management through Lyapunov optimization methods and optimizes resource allocation based on real-time feedback. The overall model can be represented by the following comprehensive formulation:

$$\begin{array}{c}\mathrm{P}1: \underset{\chi \left(t\right),f\left(t\right),p\left(t\right)}{min}\sum _{t=0}^{T-1}[\mathsf{\Delta }\left(\theta \left(t\right)\right)+V·\mathsf{\Psi }\left(t\right)]\hfill \\ \mathrm{C}1: \sum _{m\in M}\sum _{j\in N}{x}_{um}^{jl}\left(t\right)\le 1,\forall u\in U,l\in L,t\hfill \\ \mathrm{C}2: \sum _{l\in L}{f}_{u,m}^l\left(t\right)\le f_s^m\left(t\right),\forall m\in M,u\in U,t\hfill \\ \mathrm{C}3: p_u\left(t\right)\in \left[0,P_u^{max}\right],\forall u\in U,t\hfill \\ \mathrm{C}4: {\mu }_u^l\left(t\right)\ge {\mu }_{min}^l,\forall u\in U,l\in L,t\hfill \\ \mathrm{C}5: \sum _{k=1}^l{ϵ}_u^k\left(t\right)\le {\mathsf{\Gamma }}_u^{\max },\forall u\in U,l\in L,t\hfill \end{array}$$

(21)

In the formula, $\chi \left(t\right)$ represents the offloading decision set at time slot $t; f\left(t\right)$ denotes the MEC resource allocation vector; $p\left(t\right)$ indicates the user transmission power; $\mathsf{\Delta }\left(\theta \right(t\left)\right)$ is the Lyapunov drift term, representing system stability; V is a control parameter used to adjust the weight between QoE and queue stability; $\mathsf{\Psi }\left(t\right)$ is the QoE penalty function. The specific meanings of the constraints in Equation (21) are as follows: (i) Constraint $C1$ ensures that the subtasks of the same video task can only be executed locally or offloaded to one MEC server, guaranteeing a unique offloading path. (ii) Constraint $C2$ states that the total computational resources allocated by the MEC server must not exceed its current available resource limit. (iii) Constraint $C3$ requires that the transmission power of user devices must comply with the preset maximum power limit. (iv) Constraint $C4$ ensures that users receive at least the base layer data of the video stream, maintaining basic service quality. (v) Constraint $C5$ restricts the cumulative distortion of layered video transmission, ensuring overall video quality meets the standard.

As we have established the system model in Section 2, which captures the essential components and interactions within the edge computing ecosystem, the subsequent section will delve into the development of an optimization algorithm. This algorithm will be designed to effectively manage the joint offloading of tasks and allocation of resources, ensuring optimal performance and resource utilization in dynamic network environments.

3. Optimization Algorithm for Joint Offloading and Resource Allocation

Considering that a large number of variables scale linearly with the number of users, MEC servers, and sub-bands, and that real-time constraints need to be satisfied in highly dynamic mobile terminal scenarios, a low-complexity solution to the joint optimization problem with suboptimal characteristics must be designed to achieve a more competitive QoE and energy-efficiency performance while safeguarding the users’ computational needs. Since the joint optimization problem is essentially a mixed-integer nonlinear programming (MINLP) problem, the time complexity of its optimal solution search is usually exponential [36], the joint offloading and resource optimization model proposed in Equation (21) is modeled as a subproblem with a fixed binary variable $\left\{x_{um}\right\}$, which is decomposed into a subproblem with a separated objective function and several constraints [37], thus transforming the original high-complexity problem into a master problem and a set of constraints. The original high-complexity problem can be transformed into a main problem and a set of low-complexity subproblems. In summary, the unloading decision and resource allocation problems in this study are decoupled from each other. Therefore, Equation (19) can be transformed as follows:

$$\begin{array}{c}\underset{\chi }{max}{J}^{*}\left(\chi \right)\hfill \\ \mathrm{s}.\mathrm{t}. \left(\mathrm{C}1\right)-\left(\mathrm{C}3\right)\hfill \end{array}$$

(22)

3.1. Resource Allocation Issues

First assume that constraint $C1$ is satisfied, at which point the objective function can be rewritten as follows:

$$J(X,F)=\sum _{m\in M}\sum _{u\in U}{\lambda }_u\left({\beta }_u^t+{\beta }_u^e\right)-V(X,F),$$

(23)

where $V(X,F)$ is a function of X and F. The function is expressed as follows:

$$V(X,F)=\sum _{m\in M}\sum _{u\in U}{\lambda }_u\left({\displaystyle \frac{{\beta }_u^t{t}_u}{t_u^{\mathrm{local}}}} + {\displaystyle \frac{{\beta }_u^e{E}_u}{E_u^{\mathrm{local}}}}\right).$$

(24)

It can be seen that the first term of Equation (23) is constant in this study, then $V(X,F)$ corresponds to the total offloading overhead of all offloaded users, i.e., the above problem can be converted into a minimization problem of $V(X,F)$ denoted as follows:

$$V(X,F)=\sum _{m\in M}\sum _{u\in U}{\displaystyle \frac{{\theta }_u+{\vartheta }_u{p}_u}{{\log }_2\left(1+{\gamma }_{um}\right)}}+\sum _{m\in M}\sum _{u\in U}{\displaystyle \frac{{\psi }_u}{f_{um}}}$$

(25)

In the formula, ${\theta }_u={\displaystyle \frac{{\lambda }_u{\beta }_u^t{d}_u}{t_u^{\mathrm{local}}W}},{\vartheta }_u={\displaystyle \frac{{\lambda }_u{\beta }_u^e{d}_u}{E_u^{\mathrm{local}} W}},{\psi }_u={\lambda }_u{\beta }_u^t{f}_u^{\mathrm{local}}$.

Optimizing $f_{us}$ while keeping $p_u$ fixed, the computational resource allocation can be solely represented by the second term of Equation (25) as follows:

$$\begin{array}{cc}\underset{F}{\min }\hfill & \sum _{m\in M}\sum _{u\in U}{\displaystyle \frac{{\psi }_u}{f_{um}}}\hfill \\ \mathrm{s}.\mathrm{t}.\hfill & \sum _{u\in U}{f}_{um}\le f_m,\forall m\in M\hfill \\ \hfill & f_{um}>0,\forall u\in U,m\in M\hfill \end{array}$$

(26)

It can be seen that the Hessian matrix of the objective function is positive definite, and the optimization problem proposed in this paper is a convex optimization problem. According to the nature of convex optimization, the problem is solved by using the properties of Karush–Kuhn–Tucker(KKT) conditions. Then, the solution can be obtained from Equation (26):

$$f_{um}^{*}={\displaystyle \frac{f_m\sqrt{{\psi }_u}}{{\sum }_{u\in U}\sqrt{{\psi }_u}}},\forall m\in M,u\in U.$$

(27)

$$\mathsf{\Delta }\left(X,F^{*}\right)=\sum _{m\in M}{\displaystyle \frac{1}{f_m}}{\left(\sum _{u\in U}\sqrt{{\psi }_u}\right)}^2.$$

(28)

3.2. Joint Task Offloading and Resource Allocation Issues

Based on the computational resource optimization scheme given in the previous section, the task offloading joint resource allocation model can be expressed as follows:

$$\begin{array}{c}\underset{\chi }{max}\sum _{m\in M}\sum _{u\in U}{\lambda }_u\left({\beta }_u^t+{\beta }_u^e\right)-\mathsf{\Pi }\left(X,F^{*}\right)\hfill \\ \mathrm{s}.\mathrm{t}. x_{um}^{jl}\in \{0,1\},\forall u\in U,m\in M,j\in N,l\in L\hfill \\ \sum _{m\in M}\sum _{j\in N}{x}_{um}^{jl}\le 1,\forall l\in L,u\in U\hfill \\ \sum _{u\in U}{x}_{um}^{jl}\le 1,\forall l\in L,m\in M,j\in N\hfill \end{array}$$

(29)

In the formula, $\chi$ represents the decision variable matrix for resource allocation. ${\lambda }_u$ denotes the weight factor of user u, while ${\beta }_u^t$ and ${\beta }_u^e$ represent the time-sensitivity and energy-efficiency coefficients of user u. The term $\mathsf{\Pi }(X,F^{*})$ signifies the penalty cost for deviations from the optimal resource allocation configuration.

The offloading decision problem is combinatorial in nature, and a simple way to solve the problem is to use the exhaustive enumeration method to search for all task offloading decisions with possibilities, but the complexity of task offloading decisions is as high as $2^n$ when $n=M\times U\times N$. To overcome the high complexity defect of the exhaustive method, this paper adopts the HQPSO based on quantum behavioral optimization, which can find a locally optimal solution of Equation (27) in the polynomial time range. The algorithm is able to quickly approximate the global optimal solution in highly dynamic network environments through quantum bit encoding, superposition state parallel search, and a dynamic inertia weight adjustment mechanism.

Compared with traditional heuristic algorithms, HQPSO combines the parallelism of quantum computing and the group collaboration feature of particle swarm optimization: its quantum-encoding and parallel search mechanism encodes the offloaded decision variables as quantum superposition states, so that a single iteration can simultaneously explore multiple potential solution spaces, reducing the time complexity to $O(nlogn)$ [38]. Meanwhile, through the multi-objective fitness function and quantum revolving door mechanism, the algorithm can dynamically balance the optimization weights for delay, energy consumption, and hierarchical video integrity [39]. In contrast to deep reinforcement learning (DRL)-based methods, HQPSO typically incurs lower computational overhead due to its efficient quantum parallelism, which reduces the need for extensive simulation or training as required in DRL approaches. DRL methods, though powerful in dynamic environments, often demand high computational resources for model training, which can hinder scalability and real-time decision-making [40]. Moreover, while DRL models tend to prioritize convergence at the expense of interpretability, HQPSO offers a clearer decision-making process through its particle collaboration and search space exploration, allowing for better understanding and control over the optimization steps [41]. Furthermore, the convergence behavior of HQPSO is generally more predictable and faster, as the algorithm leverages the direct adjustments of inertia weights to swiftly adapt to network changes, unlike DRL, which may require a longer training phase to stabilize [42]. Therefore, HQPSO’s efficiency in both convergence and interpretability makes it particularly well suited for real-time and resource-constrained applications, where computational efficiency and transparency are critical.

In HQPSO, quantum encoding maps the offloading decision variables to quantum bits (qubits) to leverage quantum superposition for efficient optimization. Specifically, each offloading decision variable $x_{um}^{jl}\in \{0,1\}$ (indicating whether user u ’s layer l task is offloaded to MEC server m on sub-band j) is represented by a qubit $\left|q_{um}^{jl}\right.〉=\alpha |0\rangle +\beta |1\rangle$, where $\alpha$ and $\beta$ are complex amplitudes satisfying ${\left|\alpha \right|}^2+{\left|\beta \right|}^2=1$. This superposition allows a single qubit to represent both offloading choices simultaneously, enabling parallel exploration of the decision space [43]. The mapping process involves the following: (1) initializing qubits in superposition states; (2) applying quantum gates (e.g., Hadamard or rotation gates) to evolve the states based on fitness functions; (3) measuring the qubits to collapse to classical binary decisions. Mathematically, the offloading decision is decoded as $x_{um}^{jl}=\arg \max \left({\left|\alpha \right|}^2,{\left|\beta \right|}^2\right)$, ensuring the algorithm efficiently handles the combinatorial complexity by searching exponentially many states in linear time.

Figure 2 outlines HQPSO’s workflow: after initializing the quantum-bit matrix and setting pbest/gbest, the algorithm measures quantum states to obtain binary offloading decisions and then allocates resources and computes system utility under latency and energy penalties. It evaluates fitness, refreshes pbest/gbest, and updates the swarm by adjusting $\theta$ and $\omega$. If a channel-state change is detected, part of the swarm is reinitialized; otherwise, the loop returns to state measurement. The procedure repeats until the iteration budget is exhausted or convergence is achieved, at which point gbest and the associated system utility are returned.The core operations of quantum state measurement, resource allocation, and fitness evaluation are defined in Algorithm 1, which serves as the functional backbone for HQPSO’s iterative optimization process.

In hyper-dynamic environments, HQPSO demonstrates distinct advantages: quantum parallelism empowers the algorithm to achieve over $90\%$ near-optimal solutions within 5–10 iterations, satisfying the millisecond-level decision-making requirements for video streaming [44]; quantum entanglement establishes correlations among user-base station-subchannel states, maintaining $⩾95\%$ layered video transmission success rate; the dynamic subchannel allocation strategy enhances spectral efficiency by $32\%$ compared to simulated annealing while restricting computational resource fragmentation below $5\%$ [44]. Through the co-evolutionary mechanism of quantum populations, the algorithm synergistically addresses performance limitations of conventional approaches, overcoming the Greedy (GREEDY) algorithm’s myopia and simulated annealing’s stochastic oscillations in dynamic scenarios, and delivers both high-efficiency and robustness for real-time video streaming resource allocation [45]. The pseudo-code for the joint offloading decision and resource allocation algorithm based on HQPSO is provided in Algorithm 2.

Algorithm 1 Related Functions

1: Function Measure($\theta$): 2:     Generate binary offloading decision X, where $X_{im}^j=1$ if and only if ${sin}^2\left({\theta }_{im}^j\right)>\mathrm{rand}(0,1)$ 3: return X 4: Function ResourceAllocation(X): 5:     Calculate the resource allocation F according to Equation (24), where $f_{um}=f_m\ast {\displaystyle \frac{Q_d^i+V{\beta }_i^i}{\sum (Q_k^d+V{\beta }_k^i)}}$ 6: return F 7: Function $F(X,F)$: 8:     Calculate the system utility $J(X,F)$ (Equation (16)), which includes latency gain and energy consumption penalty terms. 9: return J

Algorithm 2 The joint offloading decision and resource allocation algorithm based on HQPSO

1: Input: User set U, base station set M, subcarrier set N, video layer set L, maximum iteration number $T_{\max }$, quantum swarm size $Q_{\mathrm{size}}$, dynamic inertia weights ${\omega }_{\min },{\omega }_{\max }$, fitness function $F(·)$ 2: Output: Optimal offloading decision $X^{*}$, resource allocation strategy $F^{*}$, system utility $J^{*}$ 3: Initialize the quantum particle swarm $Q=\left\{q_1,q_2,\dots ,q_{Q_{\mathrm{size}}}\right\}$, where each particle $q_i$ includes: 4: Quantum bit matrix ${\theta }_i=\{{\theta }_{i,u,m,n,l}\mid u\in U,m\in M,n\in N,l\in L\}$ (dimension $U\times M\times N\times L$), where ${\theta }_{i,u,m,n,l}\in [0,{\displaystyle \frac{\pi }{2}}]$ represents the quantum rotation angle with quantum state $|{\psi }_{i,u,m,n,l}\rangle =cos\left({\theta }_{i,u,m,n,l}\right)|0\rangle +sin\left({\theta }_{i,u,m,n,l}\right)|1\rangle$ and measurement probability $P(x_{u,m,n,l}=1)={sin}^2\left({\theta }_{i,u,m,n,l}\right)$, initialized to a $\pi /4$ superposition state 5: Historical best solution ${\mathrm{pbest}}_i=\mathrm{None}$, global best solution $\mathrm{gbest}=\mathrm{None}$, inertia weight $\omega ={\omega }_{max}$ 6: for $t=1$ to $T_{max}$ do 7:    {Quantum state observation generates candidate solutions} 8:    for each $q_i$ in Q do 9:      $X_{\mathrm{candidate}}=\mathrm{Measure}\left({\theta }_i\right)$ 10:      $F_{\mathrm{candidate}}=\mathrm{ResourceAllocation}\left(X_{\mathrm{candidate}}\right)$ 11:      $F_{\mathrm{fitness}}=F(X_{\mathrm{candidate}},F_{\mathrm{candidate}})$ 12:      {Update individual vs. global optimum} 13:      if $F_{\mathrm{fitness}}>q_i.$pbest_fitness then 14:         $q_i.\mathrm{pbest}=(X_{\mathrm{candidate}},F_{\mathrm{candidate}})$ 15:         $q_i.\mathrm{pbest\_fitness}=F_{\mathrm{fitness}}$ 16:      end if 17:      if $F_{\mathrm{fitness}}>$ gbest_fitness then 18:         $\mathrm{gbest}=(X_{\mathrm{candidate}},F_{\mathrm{candidate}})$ 19:         gbest_fitness$=F_{\mathrm{fitness}}$ 20:      end if 21:      {Quantum revolving door updating phases} 22:      $\mathsf{\Delta }\theta =\omega ·(\mathrm{pbest\_}\theta -{\theta }_i)+(1-\omega )·(\mathrm{gbest\_}\theta -{\theta }_i)$ 23:      ${\theta }_i={\theta }_i+\mathsf{\Delta }\theta$ 24:      ${\theta }_i=\mathrm{Clip}({\theta }_i,0,\pi /2)$ 25:      $\omega ={\omega }_{max}-({\omega }_{max}-{\omega }_{min})·t/T_{max}$ 26:      {Real-time disturbance response (highly dynamic scenes)} 27:      if $\mathrm{ChannelStateChanged}\left(\right)$ then 28:         $Q=\mathrm{Reinitialize}(Q,30\%)$ 29:      end if 30:    end for 31: end for 32: return$\mathrm{gbest}.X^{*}$, $\mathrm{gbest}.F^{*}$, gbest_fitness

After formulating and solving the optimization problem for joint offloading and resource allocation, it is essential to evaluate the performance of the proposed algorithm. Therefore, the next section presents simulation experiments conducted under specific scenarios to validate its effectiveness and practicality.

4. Simulation Experiment

4.1. Experimental Environment

All simulations in this study were implemented using SIMULINK R2023a on a Windows 10 system with 16 GB RAM. The experimental framework employs m-scripting language, leveraging its comprehensive libraries for linear algebra and signal processing to execute core algorithms, while utilizing SIMULINK’s modular visualization capabilities for dynamic process simulation. Note that due to confidentiality constraints involving sensitive research data, the experimental code is not currently publicly available. Should these restrictions be lifted in future, complete code resources will be provided through permanent DOI links in complementary materials or institutional repositories.

4.2. Experimental Parameters

Assume a high-speed mobile scenario system composed of multi-tier base stations, where the macro base station spacing in highway scenarios is 2 km, and the small base station spacing within subway tunnels is 200 m. The network coverage area includes seven macro base stations and fifteen small base stations. The maximum transmission power of the mobile terminal is $P_u=24\mathrm{dBm}$, the system bandwidth is $B=30\mathrm{MHz}$, and the background noise variance is ${\sigma }^2=-90\mathrm{dBm}$[29]. Users and base stations use single antennas for uplink transmission and reception, with a channel model following Rician fading ($K=6 \mathrm{dB}$, Doppler frequency shift $f_d=v·f_c/c$, and carrier frequency $f_c=3.5 \mathrm{GHz}$) [30]. In terms of computing resources, assume the edge server’s computational capability is $f_s=30 \mathrm{GHz}$, the local CPU capability of the mobile terminal is $f_u^{\mathrm{local} }= 1.5 \mathrm{GHz}$, and the energy coefficient is $\alpha =3\times {10}^{-27}$. Unless otherwise specified, the default task input data size is $d_u=800 \mathrm{kb}$, the dynamic preference parameter is ${\beta }_u^t=0.7,{\beta }_u^e=0.3$, and the safety factor is ${\lambda }_u=1.2$[34,35,36]. Under high-speed conditions, the mobile terminal follows a road-constrained random walk model (highway: linear path + lane deviation disturbance; subway: three-dimensional Brownian motion within the tunnel), with a communication latency limit of ${\tau }_{max}=50 \mathrm{ms}$. The terminal speed distribution is $[25,40] \mathrm{m}/\mathrm{s}$ for highways and $[15,30] \mathrm{m}/\mathrm{s}$ for subways.

To comprehensively validate the advantages of the proposed HQPSO algorithm in highly dynamic environments, this paper selects three typical comparative algorithms as baseline methods:

Greedy Algorithm: This method targets immediate optimality by selecting the current best allocation scheme at each step. It has low computational complexity and is suitable for fast decision-making, but it is prone to getting stuck in local optima and lacks global perspective.

Local Search (LS): This approach continuously improves solutions through neighborhood searches, which can to some extent escape the local optima of greedy algorithms. However, it has a slower convergence speed and is prone to convergence delays in highly dynamic scenarios.

Simulated Annealing (SA): By accepting suboptimal solutions with a certain probability, this method avoids local optima traps and can theoretically achieve better solutions. However, its annealing process parameters are fixed, leading to insufficient adaptability in rapidly changing highly dynamic environments.

These three methods represent typical strategies in low-complexity heuristics, neighborhood improvement, and global random search, respectively. They can fully reflect the differences in convergence, robustness, and real-time performance across different algorithm categories. Therefore, they are widely used in comparative experiments for mobile edge computing and resource allocation, serving as reasonable baselines for evaluating the effectiveness of new algorithms.

4.3. Simulation Results Analysis

Figure 3 and Figure 4 illustrate the variations in users’ average time consumption and average energy consumption with changes in preference. It can be observed that when altering the user’s preference for time ${\beta }_u^t$ (with a value range of $[0.0,0.9])$, the user’s preference for energy ${\beta }_u^e=1-{\beta }_u^t$ also changes, leading to corresponding adjustments in all users’ average time and energy consumption. As ${\beta }_u^t$ increases, the average latency decreases gradually, but this is accompanied by higher energy consumption. Additionally, as the number of users continues to rise, there is an upward trend in both the average latency and energy consumption per user. The primary reason for this phenomenon is that when a large number of users compete for system resources, the probability of each user achieving high performance during the offloading process diminishes accordingly.

Figure 5 demonstrates the convergence speed and real-time performance of the HQPSO algorithm in comparison with GREEDY, LS, and SA algorithms. The experimental results reveal distinct evolutionary trends in convergence speed among these algorithms as user count increases: GREEDY exhibits the poorest performance, followed by LS and SA algorithms, while the proposed HQPSO algorithm consistently maintains superior convergence characteristics. When user count $\le 30$, all algorithms display approximate linear growth patterns with HQPSO showing the steepest slope; in the range $0<\mathrm{user} \mathrm{count}⩽50$, LS and GREEDY begin exhibiting performance fluctuations ($\pm 2.5$ amplitude); when user count exceeds 50, HQPSO sustains steady growth while other algorithms exhibit pronounced performance degradation. The real-time superiority is further validated through HQPSO’s superior stabilization feasibility across all test scenarios.

The advantages of the HQPSO algorithm in hyper-dynamic environments stem from its hybrid quantum particle swarm optimization framework. By incorporating quantum-inspired behaviors to enhance global search capabilities, it effectively mitigates the local optima trapping issue inherent in conventional PSO while overcoming the myopic decision-making defects of GREEDY and LS algorithms. The dynamic parameter adaptation mechanism enables real-time search strategy adjustments, rapidly focusing on promising solution regions during user traffic spikes, offering greater flexibility compared to SA’s fixed cooling schedule. The performance degradation of the SA algorithm after 50 users is primarily due to its fixed cooling schedule, which causes rapid temperature decreases. As the number of users increases, the cooling rate becomes too aggressive, resulting in insufficient exploration of the solution space and leading to premature convergence on suboptimal solutions. The elite preservation strategy significantly accelerates convergence speed, with solution feasibility reaching 25 for 50 users (SA only achieves 18). Meanwhile, HQPSO eliminates LS’s convergence delay and GREEDY’s load balancing limitations, achieving superior real-time performance with reduced computational overhead, making it particularly suitable for highly dynamic scenarios.

Figure 6 illustrates the stability performance comparison among four algorithms under varying environmental dynamic strengths.The Algorithm Stability Score measures the stability of the algorithm under different environmental dynamic intensities, with higher scores indicating stronger stability. The calculation formula is as follows:

$$Stability Score={\displaystyle \frac{Stable Interval}{Fluctuation Amplitude}}$$

(30)

As demonstrated in the figure, the HQPSO algorithm significantly outperforms its counterparts: when environmental dynamic strength escalates from 0% to 100%, its stability performance metrics remain consistently within the high range of 80–90 with negligible fluctuations. In stark contrast, conventional algorithms exhibit pronounced performance degradation GREEDY plummets from 60 to 20, LS declines from 70 to 30, and SA, though relatively better, still drops from 75 to 50. This performance gap becomes particularly pronounced beyond 50% dynamic strength, where HQPSO achieves 2–4 times higher scores than competing algorithms. These results validate that HQPSO effectively addresses traditional algorithms’ performance deterioration in dynamic environments through quantum behavior optimization and dynamic parameter adaptation mechanisms. The algorithm’s unique adaptive capability establishes it as the most robust solution for hyper-dynamic scenarios.

Figure 7 demonstrates the delay performance comparison among four algorithms in hyper-dynamic environments. As illustrated, when environmental dynamic strength increases from 0% to 100%, HQPSO maintains consistently low latency within 20–40 ms with the smoothest growth curve, indicating its architectural robustness against environmental disturbances. Whereas competing algorithms exhibit dramatic fluctuations: SA surges from 20 ms to 80 ms, LS deteriorates from 30 ms to 100 ms, and GREEDY performs worst with latency skyrocketing to 120 ms. Particularly beyond the critical 50% dynamic threshold, HQPSO achieves merely 1/3 latency of GREEDY and demonstrates 50% lower delay than SA, the second-best performer. This superiority originates from HQPSO’s quantum behavior optimization mechanism, which dynamically maintains optimal path planning during abrupt environmental changes through real-time particle swarm strategy adaptation and elite preservation, while conventional algorithms suffer from fixed-parameter rigidity and local optima trapping. These results substantiate HQPSO as the optimal solution for guaranteeing ultra-low latency services in high-mobility scenarios.

Figure 8 presents a comparative analysis of energy efficiency for various algorithms under highly dynamic scenarios, highlighting the superior performance of the HQPSO algorithm. As depicted, as the environmental dynamism increases from 0% to 100%, HQPSO (red diamond) consistently maintains the highest energy efficiency (0.75–0.85 ${\mathrm{J}}^{-1}$), with a minimal decline of only 0.1 ${\mathrm{J}}^{-1}$, demonstrating the most stable and gentle curve. In contrast, other algorithms exhibit significant performance degradation: simulated annealing (blue square) drops from 0.8 ${\mathrm{J}}^{-1}$ to 0.55 ${\mathrm{J}}^{-1}$, local search (orange cross) decreases from 0.75 ${\mathrm{J}}^{-1}$ to 0.45 ${\mathrm{J}}^{-1}$, and the greedy algorithm (green circle) performs the worst, plummeting from 0.7 ${\mathrm{J}}^{-1}$ to 0.35 ${\mathrm{J}}^{-1}$. Notably, when the environmental dynamism exceeds 60%, the energy efficiency advantage of HQPSO becomes even more pronounced, achieving over twice the efficiency of the greedy algorithm and approximately 30% higher than the second-best simulated annealing algorithm. This significant advantage is attributed to the unique quantum behavior optimization mechanism of the HQPSO algorithm, which intelligently adjusts particle swarm search strategies and dynamic parameters for self-adaptation, effectively reducing unnecessary computational overhead and maintaining optimal energy utilization even in rapidly changing environments.

Figure 9 demonstrates the offloading feasibility comparison among four algorithms in hyper-dynamic environments. As illustrated, the HQPSO algorithm (red diamonds) demonstrates superior stability in energy efficiency metrics, maintaining values within 0.8–0.9 across increasing task loads with minimal fluctuations. Whereas the competing algorithms exhibit significant performance degradation: simulated annealing (blue squares) declines from 0.85 to 0.65, GREEDY (green dots) deteriorates from 0.8 to 0.55, and LS (yellow forks) performs worst, plummeting from 0.75 to 0.45. Notably, in the high-load interval (60–100), HQPSO’s energy efficiency advantage becomes more pronounced, achieving $15\%$ higher values than suboptimal SA and nearly double that of LS. This sustained high performance originates from HQPSO’s quantum behavior optimization mechanism, which dynamically adjusts particle swarm search strategies and implements intelligent resource allocation to overcome computational bottlenecks in high-load conditions while ensuring optimal offloading decisions. Moreover, the proposed HQPSO-based joint offloading and resource allocation scheme outperforms conventional GREEDY and LS algorithms in overall performance gains.

5. Conclusions

This paper proposes a resource co-optimization framework based on MEC for real-time video streaming transmission under hyper-dynamic mobile terminals (e.g., vehicular and subway scenarios), addressing challenges arising from rapidly fluctuating channel states and intense resource contention. By decomposing the problem into two subproblems—SVC-based layered video transmission optimization and dynamic edge resource scheduling—the framework leverages SVC’s hierarchical structure to partition video streams into base and enhancement layers, effectively adapting to channel variations in high-mobility environments while minimizing transmission costs and enhancing QoE. Meanwhile, through Lyapunov-based optimization, the scheme achieves dynamic task offloading and resource allocation, resolving multi-objective optimization under time-varying channel conditions while guaranteeing low latency and system stability.

The research contributions are structured as follows: First, an SVC-based hierarchical video transmission strategy is proposed, which significantly enhances the adaptability and efficiency of video streaming through layered coding and transmission mechanisms. Second, integrating Lyapunov optimization enables dynamic edge resource scheduling, effectively improving resource utilization while reducing transmission latency and energy consumption. Simulation results demonstrate that, compared to conventional approaches, the proposed framework achieves substantial improvements in resource utilization efficiency, delay performance, and energy efficiency, particularly maintaining stable video transmission in hyper-dynamic environments. Additionally, this work innovatively designs a joint offloading and resource allocation algorithm based on HQPSO, which outperforms traditional methods in convergence speed and solution quality through quantum computing’s parallel search capabilities and dynamic adaptation mechanisms, providing efficient support for real-time decision-making in highly dynamic scenarios.

The proposed framework for real-time video streaming resource optimization in highly dynamic mobile environments shows promise but faces key challenges for practical deployment. A major limitation is the computational intensity of existing HQPSO-based algorithms, which may not suit resource-constrained edge devices; future work should develop lightweight versions with reduced complexity through heuristics or fewer global optimization steps. Interoperability with evolving 5G core networks and emerging 6G systems is also critical, requiring flexible frameworks capable of dynamic adaptation across heterogeneous infrastructures. Moreover, scalability to complex network topologies with diverse node capabilities demands efficient multi-task scheduling and conflict-aware resource allocation. The robustness of HQPSO under ultra-dynamic conditions, such as sudden mobility shifts or traffic surges, must be enhanced to maintain real-time efficiency. Ultimately, future research should focus on adaptive, lightweight, and interoperable solutions that seamlessly integrate with next-generation networks, enabling efficient and resilient MEC-based video streaming.