A Lightweight Edge AI Framework for Adaptive Traffic Signal Control in Mid-Sized Philippine Cities

Abstract

Mid-sized Philippine cities commonly rely on fixed-time traffic signal plans that cannot respond to short-term, demand-driven surges, resulting in measurable idle time at stop lines, increased delay, and unnecessary emissions, while adaptive signal control has demonstrated performance benefits, many existing solutions depend on centralized infrastructure and high-bandwidth connectivity, limiting their applicability for resource-constrained local government units (LGUs). This study reports a field deployment of TrafficEZ, a lightweight edge AI signal controller that reallocates green splits locally using traffic-density approximations derived from cabinet-mounted cameras. The controller follows a macroscopic, cycle-level control abstraction consistent with Transportation System Models (TSMs) and does not rely on stationary flow–density–speed (fundamental diagram) assumptions. The system estimates queued demand and discharge efficiency on-device and updates green time each cycle without altering cycle length, intergreen intervals, or pedestrian safety timings. A quasi-experimental pre–post evaluation was conducted at three signalized intersections in El Salvador City using an existing 125 s, three-phase fixed-time plan as the baseline. Observed field results show average per-vehicle delay reductions of 18–32%, with reclaimed effective green translating into approximately 50–200 additional vehicles per hour served at the busiest approaches. Box-occupancy durations shortened, indicating reduced spillback risk, while conservative idle-time estimates imply corresponding CO₂ savings during peak periods. Because all decisions run locally within the signal cabinet, operation remained robust during backhaul interruptions and supported incremental, intersection-by-intersection deployment; per-cycle actions were logged to support auditability and governance reporting. These findings demonstrate that density-driven edge AI can deliver practical mobility, reliability, and sustainability gains for LGUs while supporting evidence-based governance and performance reporting.

1. Introduction

Urban intersections play a critical role in managing traffic flow, safety, and mobility in rapidly urbanizing cities. In many mid-sized Philippine cities, signalized intersections continue to operate under fixed-time signal plans that fail to respond to short-term fluctuations in traffic demand. This operational rigidity results in measurable idle loss time, during which vehicles remain stopped at red signals despite minimal or absent cross-traffic. Such inefficiencies contribute to unnecessary fuel consumption, increased delay, and elevated vehicular emissions. Similar operational challenges have been observed globally, motivating the development of performance-based and adaptive traffic management strategies to improve intersection efficiency and sustainability [1].

In response to these challenges, traffic signal control research has increasingly shifted toward data-driven and adaptive approaches capable of adjusting signal timings based on observed demand variability rather than static schedules [2,3]. One notable development in this transition is the emergence of Automated Traffic Signal Performance Measures (ATSPMs), which enable continuous monitoring of intersection performance and support evidence-based traffic operations through real-time feedback loops [4]. These systems improve both mobility outcomes and institutional accountability by transforming traffic observations into actionable performance indicators. Beyond controller-centric monitoring, several smart-city deployments demonstrate that heterogeneous sensing and data fusion can produce reliable network-level traffic indicators such as travel time and congestion maps. A commonly cited example is the SmartSantander city-scale testbed, which integrates fixed sensing points and mobile sensors to support services including traffic monitoring, speed/route-time observation, and congestion mapping through multi-source data integration [5]. This evidence supports the practical feasibility of augmenting vision-based intersection control with additional sensor streams when travel-time estimation is required, while motivating lightweight edge-deployable solutions for cities where such infrastructures are not yet available.

Parallel advances in edge computing and artificial intelligence (AI) have further expanded the capabilities of traffic monitoring and control systems. Recent studies demonstrate that traffic states can be estimated directly from video-based observations using computer vision models, reducing reliance on traditional loop detectors and centralized sensing infrastructure [6,7]. Despite these advances, many AI-driven traffic management solutions assume the availability of high-bandwidth communication networks, cloud-based processing, or specialized sensors. Such assumptions limit practical deployment in developing cities and local government units (LGUs), where financial, technical, and infrastructure constraints remain significant.

To address this gap, this study presents an adaptive edge-based signal control system, referred to as TrafficEZ (beta version), designed for low-cost and decentralized deployment in mid-sized Philippine cities. The proposed system integrates lightweight computer vision techniques with on-device computation to dynamically adjust signal green splits each cycle based on traffic density approximations derived from queued demand and discharge efficiency. All computations are performed locally within existing signal cabinets, enabling autonomous operation without reliance on continuous backhaul connectivity.

This research documents the installation and field evaluation of TrafficEZ at three signalized intersections in El Salvador City, Misamis Oriental—Municipal Hall, McDonald’s/Petron, and Seven-Eleven—which previously operated under a uniform 125 s, three-phase fixed-time signal plan. Using matched pre–post observations and manual loss-time measurements, this study evaluates how adaptive, density-driven signal timing affects per-vehicle delay, throughput, and reliability. Beyond operational performance, the TrafficEZ framework illustrates how edge-based AI infrastructure can support national mobility performance monitoring and compliance initiatives, including reporting mechanisms aligned with the Seal of Good Local Governance (SGLG).

In this context, this study addresses the following research questions:

• How effectively can an edge-based, camera-driven signal control system estimate real-time traffic density under mid-sized Philippine city conditions?
• To what extent does adaptive signal timing based on traffic density reduce idle loss time and improve throughput compared with fixed-time signal operation?
• How can edge-AI-enabled traffic signal control support cost-effective governance and performance reporting for LGUs with limited technical and infrastructural capacity?

This paper contributes to the literature by (1) demonstrating a deployable, energy-efficient adaptive signal control system using camera-based detection and edge computation; (2) providing a comparative field-based performance evaluation between fixed-time and adaptive signal operations under Philippine traffic conditions; and (3) framing edge AI traffic management as a practical governance innovation for resource-constrained cities.

2. Review of the Literature

2.1. Fixed-Time Signal Control and Operational Limitations

Fixed-time traffic signal control remains widely implemented in developing countries due to its low deployment cost, operational simplicity, and minimal maintenance requirements. These systems rely on predetermined cycle lengths and phase splits derived from historical traffic averages rather than real-time traffic conditions; while suitable for relatively stable demand patterns, fixed-time signal plans are inherently limited in their ability to accommodate short-term traffic fluctuations, often resulting in inefficient green time allocation and prolonged vehicle idling [8].

The primary objective of traffic signal optimization is to minimize vehicle delay, stops, and queue lengths while maximizing throughput under dynamic demand conditions [9]. However, empirical studies across multiple urban contexts have demonstrated that time-of-day and fixed scheduling strategies frequently fail to capture transient variations caused by incidents, weather conditions, pedestrian activity, and localized congestion spillbacks [2]. These operational inefficiencies contribute to increased fuel consumption and emissions, reinforcing the limitations of fixed-time control in contemporary urban traffic environments. As a result, interest in adaptive and responsive signal control strategies has intensified, particularly in cities pursuing data-driven and sustainability-oriented governance.

2.2. Adaptive and Intelligent Traffic Signal Control Systems

Adaptive traffic control systems (ATCSs) dynamically adjust signal timings based on observed traffic conditions, enabling more efficient utilization of available roadway capacity. Early large-scale implementations such as the Split Cycle Offset Optimization Technique (SCOOT) and the Sydney Coordinated Adaptive Traffic System (SCATS) demonstrated network-level improvements, including delay and throughput reductions ranging from 15% to 30% [10,11]. These systems established the foundational principle that real-time responsiveness is essential for effective signal control.

More recent research has explored intelligent signal control approaches using reinforcement learning and model-predictive control to enable decentralized and self-adaptive intersection management [6,7]; while these methods have shown promising performance in controlled environments, they often depend on dense sensor deployments, centralized computation, and high-bandwidth communication infrastructure. Such requirements present substantial barriers to deployment in resource-constrained urban settings.

Recent studies have also explored traffic signal optimization using advanced mathematical and learning-based formulations. Akopov and Beklaryan [12] proposed a parallel hybrid bi-objective genetic algorithm for improving traffic performance in large urban road networks, demonstrating substantial gains in delay reduction through global optimization of signal parameters. Similarly, multi-agent reinforcement learning frameworks have been investigated to enable coordinated signal control across intersections, where each agent learns optimal policies based on local and neighboring traffic states [13]. These approaches highlight the potential of optimization-driven and learning-based control for achieving near-optimal performance in complex networks.

However, such methods typically assume the availability of detailed traffic state representations, extensive training data, stable communications, and significant computational resources. These requirements can limit their deployability in resource-constrained environments, particularly in developing cities where traffic conditions are highly heterogeneous and operational simplicity is a critical constraint. As a result, there remains a practical need for adaptive signal control strategies that balance responsiveness and performance with computational efficiency, interpretability, and ease of deployment.

To mitigate these challenges, hybrid control approaches that combine empirical traffic density estimation with rule-based or locally optimized signal timing strategies have emerged as practical alternatives. These methods seek to balance control effectiveness with computational efficiency and infrastructure feasibility, making them more suitable for cities with limited technical capacity [14]. Collectively, this body of literature highlights the trade-off between algorithmic sophistication and real-world deployability in adaptive traffic signal control.

2.3. Transportation System Models (TSMs) and Macroscopic Control Abstraction

Traffic signal control is commonly framed within Transportation System Models (TSMs), which describe transport systems through interacting demand, supply, and control components on networks and across time scales [15]. Within TSMs, macroscopic traffic representations use aggregate state variables (e.g., flow, density, and speed) to characterize operational conditions without requiring vehicle-level trajectory modeling.

Building on this view, the present study adopts a macroscopic control abstraction for isolated intersections, where cycle-by-cycle timing decisions are driven by observable approach-level queue and discharge states derived from edge-based vision. This preserves consistency with established TSM principles while remaining deployable under mixed traffic, limited sensing infrastructure, and local operational constraints.

2.4. Computer Vision and Edge Artificial Intelligence in Traffic Management

Advances in computer vision have significantly improved the ability to detect, classify, and track vehicles in real time using video-based sensing. Deep-learning-based object detection models, including those from the YOLO family, have been widely adopted for vehicle counting and trajectory extraction due to their accuracy and computational efficiency [16]. When combined with background subtraction and contour-based analysis, these techniques enable reliable estimation of traffic density and queue formation under varying lighting and weather conditions [17].

In parallel, the adoption of edge computing—where data processing is performed locally rather than in centralized cloud servers—has gained increasing attention in intelligent transportation systems. Edge-based architectures reduce latency, improve operational reliability, and maintain functionality during network disruptions [18]. Recent systems demonstrate the feasibility of executing traffic detection and control algorithms directly on embedded processors, thereby minimizing dependence on centralized infrastructure and continuous connectivity [19,20]. These developments indicate that the integration of computer vision and edge AI provides a viable pathway for real-time traffic state estimation and decentralized signal control.

From a broader smart-city perspective, large-scale IoT testbeds such as SmartSantander, a city-scale experimental research testbed deployed in Santander, Spain, demonstrate that heterogeneous sensing and mobile probes can be fused to support city-scale travel-time and congestion assessment. These large-scale deployments reinforce the role of multi-source data integration in traffic-state understanding beyond single-sensor pipelines and provide empirical evidence that travel-time estimation becomes feasible when additional sensing modalities are available [5].

2.5. Traffic Signal Control in Developing Cities and the Philippine Context

Traffic management in developing regions presents distinct challenges, including heterogeneous vehicle compositions, informal lane discipline, pedestrian interference, and limited data infrastructure [21]. Studies conducted in Southeast Asian cities report that fixed-time signal plans perform poorly under mixed-traffic conditions, underscoring the need for adaptive and vision-based control strategies that can operate effectively without extensive sensor networks [22].

In the Philippine context, most local government units (LGUs) rely on manually timed or semi-actuated signalized intersections, often operating independently without centralized traffic management systems. National assessments by the DOST and the DPWG emphasize the importance of localized intelligent traffic solutions that align with existing infrastructure, governance structures, and funding mechanisms [23,24]. Furthermore, policy instruments such as the Seal of Good Local Governance encourage LGUs to adopt evidence-based mobility and sustainability metrics to support transparent and accountable urban management [25].

Taken together, these studies indicate a growing need for adaptive traffic signal control frameworks that integrate real-time traffic estimation, decentralized computation, and cost-effective deployment strategies. Such approaches are particularly relevant for mid-sized Philippine cities seeking to modernize traffic operations while operating under institutional and resource constraints.

3. Methodology

3.1. Method–Research Question Alignment

This study adopts a quasi-experimental pre–post field evaluation framework designed to address three research questions. RQ1 examines whether an edge-based, camera-driven system can reliably estimate real-time traffic density under mixed-traffic conditions. RQ2 evaluates the effect of density-driven adaptive signal timing on delay, throughput, and reliability relative to fixed-time operation. RQ3 assesses the practicality of edge AI signal control for local government deployment, governance reporting, and operational robustness.

3.2. Study Area and Experimental Design

The field study was conducted at three signalized intersections in El Salvador City, Misamis Oriental, Philippines: Municipal Hall (Figure 1), McDonald’s/Petron (Figure 2), and Seven-Eleven (Figure 3). These intersections are located along a mixed-traffic arterial corridor composed primarily of passenger vehicles with occasional light trucks and regular pedestrian crossings.

Prior to intervention, all sites operated under identical fixed-time signal plans. The intervention consisted of activating an adaptive edge-based signal control system, referred to as TrafficEZ, which reallocates green splits dynamically each cycle using on-device traffic density estimation. Performance under adaptive operation was compared against the incumbent fixed-time control using matched observation windows.

3.3. Baseline Fixed-Time Signal Operation

All intersections employed a three-phase signal plan with a nominal cycle length of 125 s. Phase allocations followed legacy city timing practices, with approximately 60 s allocated to mainline through and left-turn movements, 20–35 s allocated to side-street movements, and pedestrian service provided for each crosswalk. Amber and all-red intervals were held constant in accordance with local standards.

Baseline delay observations were collected while these fixed-time plans were active to establish reference performance (Figure 4).

3.4. Data Collection

3.4.1. Manual Observations for Baseline Ground Truth

Trained observers manually recorded red-phase idle time (s/vehicle) for the first platoon of approximately three to six vehicles queued at the stop bar using synchronized timers. Observations were conducted during multiple one-hour windows per day from 29 September to 3 October 2025. Data were recorded at the minute level and aggregated into hourly averages for each site and time window.

3.4.2. Sensing for Adaptive Operation

Each approach was instrumented with a fixed camera aligned to a perspective-warped region of interest covering both the queue zone (red phase) and discharge zone (green phase). Video analytics were executed entirely on-device using an embedded processor installed within the signal cabinet, ensuring autonomous operation without reliance on cloud connectivity or continuous backhaul.

3.5. Adaptive Edge-Based Signal Control

3.5.1. Density-Based Timing Model

At each signal cycle, the controller computes the next green time $t_G$ using Traffic Density Approximation (TDA) as

$$t_G=T\left(k_{R-1}\right)+T\left(k_{G-1}\right),$$

(1)

where $k_{R-1}$ captures queued demand during the preceding red phase and $k_{G-1}$ represents discharge efficiency during the preceding green phase.

Red-phase density is defined as

$$k_{R-1}={\displaystyle \frac{n}{l\times w}},$$

(2)

where n is the queued vehicle count within the queue region of interest, and l and w denote the effective queue length and lane width, respectively.

Green-phase discharge density is defined as

$$k_{G-1}={\displaystyle \frac{q_{G-1}}{v_{G-1}\times w}},$$

(3)

where $q_{G-1}$ is the mean discharge flow (veh/s) and $v_{G-1}$ is the observed discharge speed during the preceding green phase.

The mapping function $T(·)$ converts density estimates into green time allocations using calibrated gains and bounded constraints, including minimum and maximum green times, pedestrian service requirements, and fixed intergreen intervals.

3.5.2. Macroscopic Traffic State Representation and Control Abstraction

The proposed AESC framework adopts a macroscopic traffic modeling perspective. Instead of tracking individual vehicle trajectories or lane-specific movements, the controller operates on aggregated traffic state variables that summarize approach-level demand and service conditions, while vehicle-level detections are used to estimate these aggregate quantities, no vehicle-specific trajectories or lane-level control actions are used in the signal timing logic.

This modeling choice is consistent with Transportation Systems Engineering and TSMs, where traffic operations are described using aggregate variables such as flow, density, and service rates rather than microscopic vehicle interactions. Accordingly, the intersection is modeled as a discrete-time control system operating at the signal-cycle level, consistent with within-day macroscopic traffic representations in transportation systems engineering, where control actions respond to observed state evolution rather than predicted equilibrium flows [15].

In contrast to fundamental-diagram-based formulations that assume stationary relationships between flow, density, and speed, the proposed framework does not rely on equilibrium traffic conditions, nor does it assume a fixed fundamental diagram or stationary speed–density relationship. Instead, traffic states are empirically measured at each signal cycle, allowing the controller to operate under transient and congested conditions where classical steady-state assumptions may not hold.

At each signal cycle i, the traffic state is represented by the state vector

$${\mathbf{x}}_i=\left[\begin{array}{c}{k}_{R,i}\\ k_{G,i}\end{array}\right],$$

(4)

where $k_{R,i}$ denotes red-phase queue density and $k_{G,i}$ denotes effective green-phase discharge density observed during the previous cycle.

Signal control acts as a bounded state-feedback mechanism, where the green allocation for the next cycle is computed as

$$t_{G,i}=T\left(x_{i-1}\right),$$

(5)

subject to fixed safety and operational constraints, including minimum and maximum green times, intergreen intervals, phase order, and pedestrian service requirements.

From a macroscopic flow perspective, queue evolution across cycles may be expressed in discrete-time form as

$$k_{R,i+1}=k_{R,i}+a_i-s_i\left(t_{G,i}\right),$$

(6)

where $a_i$ represents aggregated arrivals during the red phase and $s_i(·)$ denotes served discharge as a function of the allocated green time.

Pedestrian movements are incorporated as fixed temporal constraints that bound the feasible signal timing space rather than as competing flow states. This formulation aligns with within-day macroscopic traffic models in Transportation System Models (TSMs), where control actions respond to observed state evolution rather than predicted equilibrium flows [15].

3.5.3. Adaptive Signal Control Algorithm

To clarify the procedural execution of the proposed macroscopic control abstraction, Algorithm 1 summarizes the per-cycle logic of the adaptive edge-based signal control (AESC) framework implemented in TrafficEZ. The algorithmic format (indentation, font style, and numbering) is retained for consistency and readability.

Algorithm 1 Adaptive Edge-Based Signal Control (AESC)

1: Initialize fixed cycle length, phase order, and safety constraints   2: Set minimum and maximum green times for each phase   3: for each signal cycle i do   4:       Acquire camera-based measurements for each approach   5:       Estimate red-phase density $k_{R,i}$ from queued vehicles   6:       Estimate green-phase discharge density $k_{G,i}$ from observed flow and speed   7:       Form macroscopic state vector $x_i={[k_{R,i},k_{G,i}]}^{\top }$   8:       Compute provisional green allocation $t_{G,i+1}=T\left(x_i\right)$   9:       Apply bounded gain and hysteresis to prevent oscillatory switching 10:       Enforce pedestrian service, intergreen, and safety constraints 11:       Apply updated green splits in the next signal cycle 12: end for

3.5.4. System Architecture and Data Flow

Figure 5 illustrates the high-level system architecture of the proposed AESC framework, highlighting the fully edge-resident execution of sensing, decision-making, and actuation.

3.5.5. Per-Cycle Adaptive Control Logic

Figure 6 depicts the per-cycle adaptive decision process used to update green splits based on observed traffic density imbalance.

3.5.6. Edge-Based Computer Vision Pipeline

Traffic density variables (n, q, v) were extracted using a lightweight computer vision pipeline consisting of homography transformation to a bird’s-eye view, YOLO-family object detection for moving and stationary vehicles, background subtraction using Gaussian mixture models, contour tracking to stabilize counts under slow-moving queues, and convex-hull trajectory analysis across the stop line to estimate headways and speeds. Morphological denoising and exponential smoothing were applied to reduce noise in flow and speed estimates.

3.5.7. Edge AI System Interface and Real-Time Outputs

To illustrate the operational realization of the proposed edge-based framework, Figure 7 and Figure 8 present screenshots from the deployed TrafficEZ system. These interfaces demonstrate how real-time detection, density estimation, and adaptive signal analytics are executed entirely on-device and exposed through a monitoring dashboard for local government operations.

The dashboard view (Figure 7) aggregates per-cycle telemetry from each intersection, including density estimates, reclaimed green time, and added discharge capacity. The live detection view (Figure 8) illustrates the computer vision pipeline that extracts vehicle counts, classifications, and queue presence, which directly feed the adaptive timing model described in the preceding subsection. Together, these figures provide visual confirmation of the system architecture, algorithm execution, and field deployability of the proposed AESC framework. Although vehicle-level detection and classification are shown for transparency and operational monitoring, the adaptive signal control logic itself does not operate at the vehicle or trajectory level. All timing decisions are derived exclusively from aggregated macroscopic state variables (queue density and discharge density), consistent with TSM abstractions, rather than from individual vehicle movements.

3.6. Calibration and Operational Safeguards

Calibration was conducted in two stages. First, geometric calibration established effective lane width w from site plans and effective queue length l through pixel-to-meter scaling in the warped image space. Second, the timing map $T(·)$ was initialized using capacity-based principles such that median-demand cycles reproduced legacy green allocations, then tuned to minimize split failures.

Operational safeguards included fixed intergreen intervals, hysteresis thresholds to prevent oscillatory phase switching, and density thresholds to suppress spurious updates under anomalous conditions.

3.7. Experimental Protocol

A before–after matched-hour observation protocol was implemented at three signalized intersections. For each site, data were collected over approximately one month, covering eight hours per day of peak and representative operating periods. This resulted in a total of approximately 240 h of observation per intersection, corresponding to about 6900 signal cycles per site under the fixed 125 s cycle length. Across all three intersections, the dataset comprises approximately 720 h of field observation, representing more than 20,000 analyzed signal cycles.

Each intersection was evaluated under fixed-time operation followed by adaptive operation using matched days and hours. During baseline periods, observers recorded red-phase idle time per vehicle. During adaptive operation, the signal cabinet logged per-cycle telemetry including queued count n, mean flow q, discharge speed v, computed green time $t_G$, and actuated splits. Intergreen intervals and pedestrian timings were held constant throughout the study.

3.8. Outcome Measures

Primary outcomes included average red-phase idle time per vehicle (s/vehicle), approach delay, and queue length. Secondary outcomes included throughput (veh/h), split utilization, and reliability indicators such as the 85th-percentile idle time and split failures. Environmental co-benefits were summarized using a conservative emissions proxy derived from avoided idle time.

3.9. Statistical Analysis

Pre–post differences were computed for each site and time window. Paired t-tests were applied for matched observations; otherwise, Welch tests with bias-corrected and accelerated bootstrap confidence intervals (10,000 resamples) were used. Effect sizes were reported as percent change and Hedges’ g. Generalized linear mixed models with random intercepts for date and hour were used to account for repeated measures.

3.10. Governance, Safety, and Reproducibility

No personally identifiable information was collected during system operation. All video processing and traffic state estimation were performed locally on edge devices, and no raw video streams were transmitted or stored. Adaptive control actions were logged at the signal-cycle level to support auditability, post hoc evaluation, and governance reporting by local government units (LGUs).

From a governance and policy perspective, the proposed adaptive edge-based signal control (AESC) framework enables incremental modernization of existing signalized intersections without requiring changes to safety-certified timing parameters, pedestrian service policies, or institutional operating procedures. The adaptive logic operates strictly within pre-approved cycle lengths, intergreen intervals, and pedestrian timings, ensuring compliance with established traffic regulations and liability frameworks.

All equations, calibration parameters, region-of-interest masks, and anonymized per-cycle telemetry logs were archived to support reproducibility and third-party validation. This design allows traffic agencies to review, verify, and justify adaptive timing decisions using transparent, interpretable macroscopic variables rather than opaque or proprietary models.

By combining local data processing, bounded control actions, and traceable decision logs, the AESC framework aligns with policy objectives related to data protection, operational accountability, and evidence-based traffic management. These characteristics make the approach suitable for deployment by resource-constrained LGUs seeking low-risk, auditable, and regulation-compatible pathways toward adaptive traffic signal control.

4. Results and Discussion

The adaptive edge-based signal control (TrafficEZ) system was activated at three signalized intersections in El Salvador City operating under a nominal 125 s, three-phase fixed-time plan. Matched-hour observations were collected before (fixed-time) and after activation (adaptive). Primary outcomes include per-vehicle red-phase idle time measured through manual observation and approach-level performance metrics derived from cabinet-logged telemetry.

4.1. RQ1: Operational Performance of Edge-Based Density Estimation

Stability and Mechanism of Adaptive Control

Under fixed-time operation, loss time accumulated when active phases failed to reflect queued demand. In contrast, TrafficEZ computes green time each cycle using a density-driven timing rule based on aggregated red-phase and green-phase traffic states (Equation (1)).

By jointly accounting for queued demand and observed discharge efficiency, the controller reallocates green time toward approaches with active queues while preserving fixed safety constraints and pedestrian service requirements, as described in Section 3.

Operational stability was evidenced by the absence of oscillatory phase switching, reduced split failures, and consistent convergence of green splits across cycles. These outcomes indicate that camera-derived density estimates were sufficiently robust for real-time control under mixed-traffic conditions. The observed stability is consistent with the macroscopic, density-based control abstraction described in Section 3, in which aggregated queue and discharge states provide a smooth and reliable basis for adaptive signal timing without reliance on vehicle-level trajectory tracking.

4.2. RQ2: Impact of Adaptive Signal Control on Per-Vehicle Delay

Red-Phase Idle Time Reduction

TrafficEZ consistently reduced per-vehicle red-phase idle time across all study sites (

Table 1. Per-vehicle red-phase idle time under fixed-time and adaptive operation.

Intersection	Baseline (s/veh)	Adaptive (s/veh)	Δ	Change (%)
Municipal Hall	14.8	11.5	−3.3	−22.5
McDonald’s/Petron	21.7	14.9	−6.8	−31.5
Seven-Eleven	8.7	7.1	−1.6	−18.0

). The largest absolute and relative reductions were observed at the McDonald’s/Petron intersection, which exhibits the greatest demand imbalance among approaches.

Reported delay values represent averages computed over thousands of signal cycles per site, with each intersection contributing approximately 6900 observed cycles during the study period.

These reductions demonstrate that adaptive, density-driven signal timing can meaningfully decrease user-experienced delay without modifying cycle length, intergreen intervals, or pedestrian service timings. The magnitude of benefit scales with approach demand asymmetry, explaining the larger reductions observed at McDonald’s/Petron relative to the more balanced Seven-Eleven intersection.

4.3. RQ2: Reclaimed Effective Green Time and Throughput Gains

4.3.1. Conversion of Idle Time into Productive Discharge

Idle time savings were translated into reclaimed effective green time on the critical approach and corresponding added vehicle discharge. Using a conservative saturation headway of 2.2 s/vehicle and an observed queued platoon of three to six vehicles (mean 4.5), the estimated reclaimed green time and throughput gains are summarized in

Table 2. Estimated reclaimed effective green time and added discharge under adaptive control.

Intersection	Reclaimed Green (s/h)	Added Discharge (veh/h)
Municipal Hall	∼216	∼98
McDonald’s/Petron	∼443	∼201
Seven-Eleven	∼102	∼46

.

Figure 9 illustrates how adaptive control converts previously unproductive red-phase idle time into productive vehicle discharge when aggregated over an hour of operation. Even modest per-vehicle delay reductions accumulate into substantial capacity gains during peak periods, particularly at intersections with pronounced demand imbalance.

4.3.2. Per-Cycle Green Split Redistribution and Queue Response

To complement the numerical performance indicators, Figure 9, Figure 10 and Figure 11 provide visual evidence of how adaptive control translates density imbalance into operational improvements; while Figure 9 summarizes the aggregate hourly recovery of effective green time and added discharge, Figure 10 illustrates the underlying per-cycle redistribution of green splits across approaches. The schematic queue illustration in Figure 11 highlights the resulting reduction in queue persistence on dominant movements. Together, these visuals reinforce the observed improvements in delay, throughput, and operational reliability reported in the preceding sections.

4.4. Comparison with Existing Adaptive Signal Control Approaches

While the present evaluation compares TrafficEZ performance against a fixed-time baseline, it is useful to contextualize the observed gains relative to established adaptive signal control systems. Network-level adaptive platforms such as SCOOT and SCATS dynamically optimize splits and offsets using centralized detection and continuous communications, and have reported average delay reductions in the range of 15–30% under favorable operating conditions. However, these systems typically require extensive detector infrastructure, stable backhaul connectivity, and specialized maintenance capability, which can limit their deployability in mid-sized and resource-constrained cities.

In contrast, TrafficEZ operates as a cabinet-resident, rule-based adaptive layer that reallocates green time using locally observed density imbalance while preserving the original cycle length, phase order, and pedestrian safety timings. Although the proposed system does not perform network-wide coordination or predictive optimization, the magnitude of delay reduction observed at the study sites (18–32% reduction in red-phase idle time) is comparable to reported improvements from centralized adaptive systems, particularly at intersections exhibiting strong demand asymmetry. Unlike global optimization or learning-based controllers that rely on extensive state representations, training data, and communications infrastructure [12,13], the proposed AESC prioritizes bounded, interpretable adaptation suitable for legacy signal hardware and mixed-traffic operating conditions.

From an operational perspective, TrafficEZ occupies a middle ground between fixed-time control and fully centralized adaptive systems: it delivers measurable responsiveness to demand variation without the cost, institutional complexity, or communications dependence associated with large-scale adaptive platforms. This positioning makes the approach particularly suitable for incremental modernization of existing intersections in developing-city contexts.

4.5. RQ3: Reliability, Pedestrian Service, and Governance-Relevant Outcomes

4.5.1. Reliability and Pedestrian Compliance

Pedestrian walk and flashing-do-not-walk intervals were preserved at all sites. Call-to-walk service was delivered within two cycles, and increased split utilization occurred primarily on critical vehicular approaches rather than overserving minor movements. Reliability improved across sites, with the most pronounced contraction in upper-tail delay (approximately the 85th percentile) observed at McDonald’s/Petron, indicating reduced spillback risk and fewer extreme waits.

4.5.2. Environmental Proxy and Governance Implications

Idle fuel consumption and CO₂ reduction were evaluated using a simplified, proxy-based approach based on avoided red-phase idle time. The resulting values should be interpreted as approximate estimates intended to illustrate the relative environmental benefit of adaptive operation, rather than precise emissions inventories.

Using conservative idle fuel and CO₂ conversion factors appropriate for mixed vehicle fleets, avoided idle time corresponds to estimated emissions reductions of approximately 0.05–0.10 kg/h at Municipal Hall, 0.10–0.13 kg/h at McDonald’s/Petron, and 0.02–0.03 kg/h at Seven-Eleven during peak periods. Emission estimates were derived using conservative idle emission factors representative of light-duty vehicles, which dominate the observed traffic stream at all study sites.

These results demonstrate how localized delay reductions translate into measurable environmental co-benefits and are sufficient for comparative evaluation and planning-level assessment, which is the intended scope of this study.

4.6. Implications for Mixed Traffic and Southeast Asian Operating Conditions

Traffic conditions in Philippine and Southeast Asian cities differ substantially from those assumed in lane-based or trajectory-driven adaptive signal control systems. Intersections typically operate under mixed traffic composed of motorcycles, tricycles, jeepneys, private vehicles, and buses, often with weak lane discipline, informal queue formation, and frequent lateral movements. Under such conditions, lane-precise detection and vehicle-level trajectory tracking can become unreliable and computationally expensive.

The adaptive edge-based signal control (AESC) framework is intentionally designed around density aggregation rather than lane-specific vehicle tracking. By integrating detector occupancy and vehicle presence over rolling time windows, the system remains robust to short-term lateral movements, heterogeneous vehicle behavior, and partial violations of lane discipline. This aggregation-based approach reduces sensitivity to individual detection errors and is therefore well suited to mixed-traffic environments common in Philippine and Southeast Asian cities.

Variable lighting and weather conditions—such as glare, shadows, rainfall, and partial occlusion—also affect camera-based sensing in tropical regions. To mitigate these effects, the AESC employs conservative density thresholds, temporal smoothing, and bounded split adjustments rather than instantaneous per-frame decisions. As a result, transient sensing noise does not translate into abrupt or unsafe timing changes at the controller level.

These design choices align with the operational realities of Philippine local government units, where detector maintenance resources are limited, connectivity may be intermittent, and safety-certified timing parameters must be preserved. Compared with learning-based or lane-precise adaptive controllers, the proposed density-driven, rule-bounded approach offers a more deployable, interpretable, and maintainable pathway for adaptive signal control in Southeast Asian urban corridors.

4.7. Limitations and Practical Implications

Manual red-phase idle-time extraction was used due to local instrumentation constraints. To reduce observer bias, observations were conducted by trained traffic enforcers following a fixed measurement protocol consistent with standard traffic survey practices used by the MMDA. Manual observation remains commonly used for signal timing and validation in Philippine cities, and human annotations are frequently employed as ground truth for camera-based traffic studies when controller high-resolution logs are unavailable [26,27].

Although idle times were manually recorded, the large number of observed signal cycles per site (approximately 6900) mitigates individual measurement noise, and the before–after matched-hour design further reduces systematic bias.

Camera-based sensing performance under poor lighting and adverse weather is an important operational consideration; while the field evaluation primarily covers daytime conditions, the deployed TrafficEZ system already incorporates robustness measures, including day-to-night style transfer and GAN-based data augmentation during model training. At runtime, conservative density thresholds, temporal smoothing, and bounded green-split adjustments prevent transient sensing noise from causing abrupt or unsafe timing changes.

The environmental analysis is based on avoided idle time multiplied by standard emission coefficients and therefore represents an order-of-magnitude estimate intended for comparative assessment rather than a detailed emissions inventory.

This study does not perform a direct experimental comparison with other adaptive systems such as SCATS, SCOOT, or learning-based controllers, which require infrastructure and institutional conditions not present at the study sites. Instead, the objective is to demonstrate the achievable performance gains of a low-cost, edge-based adaptive strategy relative to fixed-time operation.

From a policy perspective, the proposed AESC framework is directly applicable to small and mid-sized cities with isolated or semi-coordinated signalized intersections and basic camera infrastructure. Adaptive control operates strictly within pre-approved cycle lengths, pedestrian timings, and safety constraints, enabling incremental deployment by local government units without modifying existing regulatory or operational frameworks.

4.8. Measurement Validity

Several limitations warrant consideration. Baseline delay measurements relied on manual observation rather than continuous controller logs, and throughput gains were estimated using standard saturation headway assumptions and observed queue lengths. Intersections with heavier or more persistent queues would be expected to yield proportionally larger benefits under adaptive control.

Despite these limitations, observed delay reductions of 18–32% and added discharge of approximately 50–200 vehicles/h at the busiest approaches demonstrate that edge-based, density-driven signal control can deliver meaningful mobility, reliability, and sustainability benefits for resource-constrained Philippine cities without altering established safety or pedestrian policies.

5. Conclusions and Recommendations

5.1. Conclusions

This study investigated the effectiveness of an adaptive, edge-based traffic signal control system for improving intersection performance in mid-sized Philippine cities. By integrating camera-based traffic sensing with on-device computation, the proposed TrafficEZ system dynamically reallocates signal green splits each cycle based on real-time traffic density estimates, without reliance on centralized infrastructure or high-bandwidth connectivity.

In response to RQ1, the results demonstrate that edge-based, camera-driven density estimation can reliably support real-time signal control under mixed-traffic conditions. The stability of per-cycle signal adjustments, reduced split failures, and consistent convergence of green allocations indicate that the derived density metrics are sufficiently robust for operational deployment at signalized intersections.

Addressing RQ2, the quasi-experimental field evaluation shows that density-driven adaptive control reduced per-vehicle red-phase idle time by 18–32% across all study sites compared with fixed-time operation. Reclaimed idle time was converted into productive service on critical approaches, yielding estimated throughput gains of approximately 50–200 vehicles per hour at the busiest intersections. These improvements were achieved without modifying cycle length, intergreen intervals, or pedestrian safety timings, highlighting the operational conservatism of the approach.

With respect to RQ3, the findings indicate that edge-AI-enabled signal control offers a practical and governance-compatible solution for local government units with limited technical capacity. All control decisions were executed locally within the signal cabinet, ensuring robustness during network interruptions and enabling incremental, intersection-by-intersection deployment. Logged per-cycle telemetry provides auditable performance evidence that can support data-driven traffic management and compliance with institutional frameworks such as the Seal of Good Local Governance.

Overall, this study demonstrates that density-driven edge AI can deliver meaningful mobility, reliability, and sustainability benefits at signalized intersections while remaining compatible with existing infrastructure, safety policies, and governance requirements in resource-constrained urban contexts.

5.2. Recommendations and Future Work

Based on the findings of this study, several directions for further development and application are recommended. First, future deployments should integrate continuous Automated Traffic Signal Performance Measures (ATSPMs) to complement manual observations and enable long-term monitoring of delay, split utilization, and reliability. This would strengthen performance evaluation and support proactive traffic operations.

Second, corridor-level coordination among adjacent signalized intersections should be explored by extending the edge-based framework to exchange limited state information between controllers. Such coordination has the potential to reduce platoon fragmentation and further enhance travel time reliability along arterial corridors.

Third, more comprehensive safety surrogate analyses—such as post-encroachment time (PET) and time-to-collision (TTC)—should be conducted using the available vehicle trajectory data to quantitatively assess safety impacts beyond operational efficiency.

Finally, broader pilot implementations across diverse urban contexts are recommended to examine scalability, seasonal variability, and long-term governance outcomes. Embedding edge-based adaptive signal control within LGU planning and reporting processes can support evidence-based mobility management while advancing sustainability objectives in developing cities.