Next Article in Journal
Kinodynamic Model-Based UAV Trajectory Optimization for Wireless Communication Support of Internet of Vehicles in Smart Cities
Previous Article in Journal
Robust Optimization Models for Planning Drone Swarm Missions
Previous Article in Special Issue
An All-Time Detection Algorithm for UAV Images in Urban Low Altitude
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Drones
Volume 8
Issue 10
10.3390/drones8100573
drones-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

UAV Hunter: A Net-Capturing UAV System with Improved Detection and Tracking Methods for Anti-UAV Defense

by
Tao Zhang
,
Ruitao Lu
*,
Xiaogang Yang
,
Xueli Xie
,
Jiwei Fan
and
Bin Tang
Department of Missile Engineering, Rocket Force University of Engineering, Xi’an 710025, China
*
Author to whom correspondence should be addressed.
Drones 2024, 8(10), 573; https://doi.org/10.3390/drones8100573
Submission received: 7 September 2024 / Revised: 6 October 2024 / Accepted: 9 October 2024 / Published: 11 October 2024
(This article belongs to the Special Issue Advances in Detection and Tracking Applications for Drones and UAM Systems)
Browse Figures
Versions Notes

Abstract

:
The abuse of UAVs poses a potential risk to social security, necessitating the investigation of anti-UAV methods to safeguard critical areas. However, the existing UAV countermeasures face challenges such as high environmental impact, restricted spatial deployment, and low cost-effectiveness. To address these limitations, we developed a novel anti-UAV system known as UAV Hunter, which adopts an airborne tether-net capture device with visual aids to counter unauthorized UAVs. This system employs an “Anti-UAV with UAV” scheme, comprising a ground control station and a net-capturing UAV. The operator utilizes the ground control station to determine the mission area and flight path and then controls the flight of the net-capturing UAV. During flight, the net-capturing UAV leverages its dual-mode sensor to continuously monitor the target area. Simultaneously, the onboard computer executes a UAV detection and tracking algorithm to search for unauthorized UAVs in real time. The results are relayed to the operator in real time, facilitating precise adjustments for the net-capturing UAV to launch the rope net accurately. The system successfully realizes the functions of dual-mode real-time detection and tracking, precise net capture, and efficient integrated control. Compared with existing methods, the developed system exhibits accurate recognition, rapid action, diverse application scenarios, and an enhanced human–machine interaction experience. Test results in the open environment further validate the feasibility and functional integrity of the system, demonstrating its capability to effectively capture low-altitude unauthorized UAVs.

1. Introduction

With the gradual opening up of low-altitude airspace, the application of unmanned aerial vehicles (UAVs) has witnessed rapid growth. However, due to inadequate regulations and poor countermeasures, the phenomenon of “black flight” for UAVs occurs frequently, posing a serious potential threat to airports, communities, and other critical locations [1,2]. Thus, it is urgent to conduct countermeasure research against these unauthorized UAVs.
Currently, anti-UAV methods can be categorized into two main types: soft destroy methods and hard destroy methods. Soft destroy methods functionally disrupt non-cooperating UAVs at the communications level by jamming or spoofing their radio or GPS signals. These methods include navigation deception, electromagnetic suppression, and communication link seizure [3,4]. While soft destroy methods have the advantages of being cost-effective and sustainable, their countermeasure range is limited, and they are susceptible to environmental factors. On the other hand, hard destroy methods aim to physically damage unauthorized UAVs. These methods include dense artillery interception, high-energy microwave attacks, and laser attacks [5,6]. Although these methods possess mature technology and highly effective damage capabilities, they also entail high costs and are likely to cause collateral damage to various installations and living organisms, making them unsuitable for deployment in densely populated and built-up areas.
In recent years, the development of UAVs and artificial intelligence (AI) technology has provided new options for countering unauthorized UAVs [7,8,9,10,11,12,13,14,15,16]. One such option is to use UAVs equipped with countermeasure equipment to counter UAVs at close range, known as “Anti-UAVs with UAV” schemes. Specifically, Xie et al. proposed a spatio-temporal centric Siamese network that offers an encouraging solution for anti-UAVs [1]. Aiming at the confrontation of low, small, and slow non-cooperating UAVs, Xiang et al. proposed the technical scheme of anti-UAV flying net launch, established a default beam unit model and a mechanical model of a flying net, and effectively intercepted UAVs [17]. These approaches offer several advantages over traditional countermeasures, including the ability to break through spatial limitations and achieve more efficient results. Nevertheless, due to the limited payload capacity of UAVs, most mainstream countermeasure equipment is not suitable for direct mounting on UAVs. Fortunately, tether-net capture devices offer a promising solution to this problem. Compared to other existing apparatus, a net-capturing device has the advantages of being compact, cost-effective, and causing no additional damage. These features render them well-suited for utilization in UAV-based anti-UAV systems.
However, several critical issues must be addressed in the development of an effective anti-UAV system utilizing airborne net gun devices. Firstly, we need to deploy an efficient UAV-to-UAV detection and tracking algorithm that assists the operator to cast the rope net precisely. This algorithm should be capable of accurately recognizing and tracking unauthorized UAVs in various dynamic scenarios, running in real time on an onboard computer. Secondly, we need to implement a remotely controlled tether-net capture device carried by the UAV, which is necessary for long-distance human–machine interaction. Thirdly, it is imperative to integrate UAV control, sensor pod control, and tether-net capture control into a unified unit, allowing live images to be displayed and devices to be operated simultaneously from a single remote control.
Based on the above analysis, we designed a novel anti-UAV system based on a vision-assisted airborne tether-net capture device, called UAV Hunter. This system mainly includes a ground control station and a net-capturing UAV. The ground control station is responsible for planning the flight path of the net-capturing UAV. When an unauthorized UAV is detected, the ground control station automatically plans the flight path and sends commands to control the net-capturing UAV to patrol the target airspace. The net-capturing UAV is a six-rotor UAV equipped with a visual assistance system and an airborne tether-net capture device, designed for close-in interception of unauthorized UAVs. The visual assistance system is composed of a dual-mode sensor pod and an onboard computer. During patrol, the dual-mode sensor pod is used to survey the target airspace all day. The captured images are processed by an onboard computer in real time and transmitted back to the ground remote control, assisting the operator in precisely locating the unauthorized UAV. Simultaneously, the operator can remotely launch a rope net to counteract the unauthorized UAV. Compared to traditional methods, the designed system has successfully achieved dual-mode real-time detection, remote net capture, and one-machine control, offering advantages such as accurate recognition, swift action, diverse application scenarios, and heightened intelligence. The primary contributions of this paper are as follows.
(1) We developed a novel anti-UAV system based on a vision-assisted airborne tether-net capture device, named the UAV Hunter. This system adopts the scheme of “Anti-UAVs with UAV”, utilizing a net-capturing UAV to counter unauthorized UAVs in close proximity. The entire process is remotely controlled over long distances and can be completed by a single operator, enhancing safety and effectiveness while also overcoming the spatial limitations of the traditional anti-UAV methods.
(2) We implemented a remotely controlled tether-net capture device with visual assistance, which is mounted on the six-rotor UAV along with a dual-mode sensor pod and an onboard computer. With the assistance of intelligent UAV detection and tracking, the operators can remotely activate the tether-net capture device from the ground in time to accurately counter unauthorized UAVs.
(3) Based on PP-YOLO tiny and the improved ByteTrack, we realized a dual-mode UAV detection and tracking algorithm, which is deployed on the onboard computer. This algorithm enables continuous and accurate detection and tracking of unauthorized UAVs from the perspective of the net-capturing UAV. Accelerated by TensorRT, the algorithm runs at 42.5 FPS with an mAP of 95.2% on the onboard computer, meeting operators’ requirements for real-time and accuracy.
(4) To further enhance the human–machine interaction experience, we integrated an image display function and multiple control functions into a single remote control, which enables the operator to quickly obtain auxiliary positioning information and accurately launch the rope net remotely to intercept targets.

2. Related Work

2.1. Anti-UAV Methods

In response to the growing threat of UAV misuse, numerous anti-UAV methods have been proposed. These methods can be broadly categorized into three groups: interference blocking, direct destruction, and net capture. Interference blocking is a soft damage method that involves terminating the communication link between a UAV and its flight control platform, thereby forcing an unmanned landing. In contrast, direct destruction methods belong to the hard damage category and are designed to strike UAVs directly on a physical level. The third approach is the net capture class, where a flexible rope net is employed to intercept or capture unauthorized UAVs [7,8].
Among these various countermeasure methods, the use of rope-net capture has gained prominence as a novel approach owing to its high adaptability, minimal requirements, and affordability, thereby facilitating its expeditious advancement. Wei et al. [18] established a finite element model of a mass–truss element rope net and conducted simulation experiments in ABAQUS software. Wang et al. [19] was inspired by the idea of using high-altitude area folding nets to capture the target square and developed a portable anti-UAV intelligent intercepting network system. Su et al. [20] invented an anti-UAV projectile soft damage device using the principle of high-pressure gas ejection to achieve UAV countermeasures. The existing research on UAV net capture systems mainly focuses on theoretical feasibility analysis, which deserves further research for practical application and development.

2.2. UAV Detection and Tracking Technology

UAV detection methods primarily include radar, radio frequency, audio, and visual detection [21]. The radar detection method is capable of determining the spatial position of a target [22]. However, the deployment of radars on UAVs can be constrained by factors such as size, weight, and power consumption, which can limit their use in certain scenarios. The audio detection method is significantly affected by environmental noise, which makes it challenging to detect UAVs in noisy environments [23]. The radio detection method requires high-sensitivity transmitting and receiving equipment [24]. With the outstanding performance of deep learning in various computer vision tasks and the widespread use of high-resolution cameras, it is now feasible to detect low-altitude UAVs based on visual detection technology.
The popular visual detection algorithms can be broadly categorized into one-stage and two-stage algorithms. One-stage algorithms, known for their compact and efficient model structures, are particularly suitable for real-time detection tasks. Some prominent one-stage detection algorithms include the YOLO series [25,26,27,28], the SSD series [29,30], Center-Net [31], and others. Aiming at the detection of low-altitude UAVs, Zhou et al. [32,33] proposed an approach based on the YOLO framework to address issues such as object interference, motion blur, and complex scenes in detecting UAVs in challenging environments. Jing et al. [34] performed spatial feature fusion based on Center-Net, and Zhang et al. [35] optimized its label coding strategy and bounding box regression, which improved UAV detection accuracy. A feature pyramid and sample adjustment factor were constructed based on the SSD network, which is effective for the scale transformation of UAVs [36]. Zhou [37] further proposed the SSD-MobileNetV2 network for feature pyramid fusion and network model scaling. Meng et al. [38] adopted MobileNetV3 and knowledge distillation to achieve network lightweight so as to improve detection performance. Long et al. [39] proposed PP-YOLO and provided a comprehensive set of quantization tools for easy deployment in edge computing devices. Based on this, through experimental verification, we chose PP-YOLO tiny as the airborne target detection algorithm to locate unauthorized UAVs in real time and accurately.
The high-speed and low-altitude maneuvers of UAVs often result in tracking algorithms being affected by confounding factors, such as complex backgrounds, noise, and motion blur [40,41,42]. These challenges significantly impede the accuracy and reliability of target tracking [43,44,45,46]. Wu et al. [47] evaluated various UAV tracking algorithms on four datasets and elaborated on UAV target tracking techniques in video. Osco et al. [48] summarized UAV tracking algorithms in different fields in the literature, such as environmental monitoring, urban governance, and agricultural management. Srivastava et al. [49] presented recent advances in 2D target detection from the UAV view, focusing on the differences between typical and UAV views. Cazzato et al. [50] reviewed UAV tracking algorithms based on correlation filtering and a tracking algorithm based on deep learning. Li et al. [51] focused on the broad application of the YOLO series in UAV tracking algorithms. Zhang et al. [52] proposed a simple and effective multi-target tracking algorithm, ByteTrack, which effectively solves the tracking loss problem due to occlusion by using low-scoring frames for secondary matching. ByteTrack discards the complex ReID module, allowing it to be utilized directly for target association without training. To achieve real-time unauthorized UAV tracking with a low computational burden, we selected the tracking algorithm that integrates PP-YOLO tiny and Byte data association after several experiments.

3. Overview of the Proposed Method

To tackle the problem of unauthorized UAV incursions, we developed the UAV Hunter, an anti-UAV system with a vision-assisted airborne tether-net capture device, as shown in Figure 1. The main framework of this system is depicted in Figure 1, which is primarily composed of two components: the ground side and the key side. The ground side is responsible for the flight trajectory planning and remote control of the net-capturing UAV, while the sky side employs the net-capturing UAV to execute missions related to unlicensed UAV detection, tracking, and countermeasures.
Initially, the ground control station determines the target search area and plans the flight trajectory of the net-capturing UAV. Subsequently, the net-capturing UAV with a dual-mode sensor pod patrols the mission airspace. Concurrently, the onboard computer processes sensor images in real time using the detection and tracking algorithm and transmits the results to the ground remote controller. With the aid of the returned results, the ground operator can precisely identify the target and remotely control the airborne tether-net capture device to launch the rope net to apprehend unauthorized UAVs.

3.1. Hardware Framework

The hardware framework of the proposed UAV Hunter system, as depicted in Figure 1, primarily consists of the ground computer, the remote control, and the net-capturing UAV. The ground computer serves as the ground station for planning flight trajectories and monitoring the state parameters of the net-capturing UAV. The remote controller acts as a unifying tool for controlling multiple devices, including UAV flight control, pod angle adjustment, and airborne tether-net capture triggering, while also displaying return results. The net-capturing UAV is the key component of this system, responsible for detecting and capturing unauthorized UAVs. The design of the net-capturing UAV is crucial to the system’s functionality, so we provide a detailed overview of its hardware structure.

3.1.1. Net-Capturing UAV

The hardware composition of the net-catching UAV, which utilizes a six-rotor UAV as the flight platform and is equipped with a visual assistance system and an airborne net capture device, is depicted in Figure 2. The component models of the system are shown in Table 1.
We assembled a six-rotor UAV using open-source modules. The six-rotor UAV, in contrast to four-rotor UAVs, boasts an enhanced power system due to the addition of two new rotor powers. This feature enhances the UAV’s payload capacity, flight stability, and anti-jamming capabilities, making it well-suited for anti-UAV missions. The open-source V5+ flight controller was adopted as the primary control unit of the six-rotor UAV. The controller is equipped with a built-in three-axis accelerometer, a three-axis magnetometer, and a three-axis gyroscope, in addition to external modules, such as GPS, barometer, and data transfer modules. As shown in Figure 3, these components collectively facilitate precise flight control, information communication, and state acquisition and filtering of the UAV.
The visual assistance system employs an airborne computer as the processing center and utilizes a dual-mode sensor pod for target detection, which provides the operators with precise target location information. We employed the SHD10T3 visible-IR zoomable pod for continuously acquiring images of the target airspace throughout the day. SHD10T3 is a highly accurate two-axis gimbal pod with a 10× zoom capability and supports the HDMI interface output. This device boasts several advantages, including its compact size, lightweight design, and low power consumption. The visible imaging resolution is 1920 × 1080 pixels, and the infrared imaging resolution is 640 × 512 pixels. The built-in shock-absorbing ball can improve the stability of image acquisition.
The onboard computer is used to process the images captured by the SHD10T3 pod by running UAV detection and tracking algorithms. We chose the NVIDIA Jetson Xavier NX as the onboard computer because of its high-performance GPU and rich peripheral interfaces. The device is compact and boasts 21 TOPS of computing power, consuming a mere 15 watts of power. It is capable of running deep learning algorithms efficiently to process sensor images, enabling accurate positioning of target UAVs.

3.1.2. Airborne Tether-Net Capture Device

In order to implement UAV countermeasures, we developed a remote-controlled tether-net capture system as the airborne net capture device. The overall configuration of the airborne tether-net capture device, which is mainly composed of a bracket, an electronic control unit, and a net-capturing device, is depicted in Figure 4.
The bracket is composed of a lightweight and robust carbon fiber material, weighing approximately 0.5 kg and measuring 113 mm × 65 mm × 140 mm in volume. Its primary function is to firmly attach the tether-net capture device to the six-rotor UAV.
The electronic control unit (ECU) is compact and operates at a power supply voltage of 8 to 12 V, which is supplied by a 3S aviation model lithium battery. The ECU is activated by a PWM electric control signal.
The tether-net capture device, weighing approximately 1 kg and measuring about 250 mm in length, with an average diameter of around 70 mm, is composed of a main body, a netting head, and a power unit. The main body is used to connect the net capture head and the power unit. The net capture head houses a 3 m × 3 m woven net with a mesh count of 15, whose deployment is managed by eight mounted weights. The power unit employs a pneumatic design, utilizing high-pressure gas to propel the rope net. This approach is not only safer and more reliable than gunpowder, but also immune to the effects of expiration or damp.
The airborne tether-net capture device is mounted at a fixed angle of 45°. After several tests, it was found that the rope net unfolds around the third meter and exhibits a significant trajectory deviation around the eighth meter. As a result, its effective capture range is about 3 to 8 m. To facilitate visually assisted estimation, we adjusted the angle of the pod to 45° as well. In this case, once the target is situated at the center of the sensor image, the operator can instantly activate the tether-net capture device to launch the rope net, thereby achieving accurate capture of the target.

3.2. Software Framework

The system’s software is categorized into two main components: ground-side software and sky-side software. The ground-side software is the UAV navigation and control software, which is deployed at the ground control station. The sky-side software is primarily the UAV detection and tracking software running on the onboard computer.

3.2.1. UAV Navigation and Control Software

Net-capturing UAVs can be controlled by remote control devices and ground control stations (GCSs). A GCS is equipped with UAV navigation and control software, which is utilized for automatic UAV patrol control. This software provides functions such as trajectory planning, operation control, status monitoring, and data display. Specifically, the status monitoring function acquires real-time data on a UAV’s current state, including position coordinates, attitude information, etc., and presents this information through virtual meters or other controls. Additionally, the GCS provides a user-friendly interface that allows the operator to enter commands, monitor the status of a UAV, and display its location on a map. The operation control functions include takeoff, return, and landing control; flight control; and data link control. Simultaneously, the GCS can display and record flight status parameters, orbit information, and load information. After the operator demarcates the patrol area, the GCS automatically plans a flight trajectory. This plan is then transmitted via a digital transmission module to the net-capturing UAV, which autonomously navigates along the pre-arranged path and executes the patrol mission.

3.2.2. UAV Detection and Tracking Software

In order to realize accurate and real-time UAV detection and tracking on the onboard computer, we adopted the lightweight PP-YOLO tiny and improved ByteTrack to develop software for intelligent processing of the dual-mode sensor images. Specifically, we employed PP-YOLO tiny as the UAV detection algorithm to efficiently identify unauthorized UAVs and then adopted the improved ByteTrack to correlate the detection results of the front and back frames to obtain the motion trajectories of unauthorized UAVs. Figure 5 shows a framework diagram of the detection and tracking algorithm.
PP-YOLO tiny is an ultra-lightweight object detection algorithm and an improvement on PP-YOLO. It follows the main framework of PP-YOLO and retains the optimization methods, such as SPP [53], IOU Loss [54], mixup, etc. It further considers the limitations of mobile devices and employs various optimization strategies to achieve efficient performance on mobile devices. The main framework of PP-YOLO tiny, which mainly includes the backbone, neck, and head, is shown in Figure 6. To effectively extract deep features, MobileNet V3 [55], which is a lightweight CNN developed through Neural Architecture Search (NAS), was selected as the backbone. The feature pyramid network (FPN) is employed as the neck to fuse the deep semantic information and shallow detail features. The structure of the head is as simple as YOLOv3, using 1 × 1 convolution to adjust the number of channels. By contrast, PP-YOLO tiny introduces depth-wise separable convolution into the detection head. This operation reduces the parameters and computational cost, making it more suitable for deployment on mobile devices. Considering the hardware differences between mobile devices and servers, PP-YOLO tiny removes optimization strategies, such as IOU aware and matrix NMS, which have a negative effect on the model size and speed.
In terms of detection algorithm selection, we selected the existing advanced detection algorithms to accomplish the task of UAV target detection from a UAV perspective. By comparing several advanced, deployable, and lightweight algorithms, and considering both accuracy and real-time performance in the task, we ultimately chose the PP-YOLO tiny scheme. This choice was based not only on its outstanding detection performance but also on its potential for practical application in UAV detection tasks. To realize fast and accurate dual-mode UAV detection, we employed the prepared dual-mode UAV detection dataset to train PP-YOLO tiny on a high-performance workstation located on the ground. During the training, to improve the recall rate for small-sized objects, we set the size of the input images to 416 × 416 and implemented the anchor box clustering approach, which obtains the priori anchor parameters that are adapted to the scale distribution of the objects. During the inference, in order to achieve real-time UAV detection on the onboard computer, we utilized PaddleSlim to quantize and compress the model and employed the Paddle Inference with activated TensorRT to accelerate the inference process. Following the dynamic offline quantization of the PaddleSlim, the model weights can be saved as int8 data, thus compressing the model volume to 1.3 M. The quantized and accelerated model can achieve an average inference speed of 40.7 FPS on the NVIDIA Jetson Xavier NX.
Considering the potential inadaptability of linear Kalman filtering to the instability of UAVs during flight, we introduced innovative improvements to the tracking method. Additionally, ByteTrack provides a simple, efficient, and generalized data association method. This method eliminates the REID module and necessitates no offline training of model parameters, making it particularly suitable for mobile devices. Therefore, we designed the improved ByteTrack method specifically for tracking unauthorized UAVs. This involved refining the ByteTrack state description vector and developing a Kalman filter model that is better suited to handling the non-uniform motion of UAVs, among other enhancements. Many tracking algorithms include Deepsort and ByteTrack, using an eight-dimensional vector description of the target’s movement: x = u , v , r , h , u ˙ , v ˙ , r ˙ , h ˙ T . This state description vector is a widely used format at present. However, we found that the bounding boxes predicted using this state vector format did not always fit the targets well, especially under conditions where the detector performance was suboptimal. In this paper, we propose an improvement to the ByteTrack state description vector by directly predicting the width and height. The modified state vector is represented as x = u , v , w , h , u ˙ , v ˙ , w ˙ , h ˙ T , where u , v is the central point coordinate of the prediction box; w and h are the width and height, respectively; and the remaining four terms refer to their corresponding rates of change. Additionally, in conventional Kalman filter frameworks, when assuming uniform linear motion for UAVs, the state vector typically solely comprises position and velocity components. However, recognizing that UAVs do not maintain constant speed during flight, we incorporated acceleration as an additional state variable to accommodate acceleration and deceleration scenarios. By introducing acceleration into the state vector and adjusting the state transition matrix and observation matrix accordingly, we devised a Kalman filter model that is better suited to handling the non-uniform motion of UAVs. The state transition matrix interrelates the three state variables—position, velocity, and acceleration—thereby fulfilling the tracking requirements for moving targets.
Given p = u , v , r , h , s = u ˙ , v ˙ , r ˙ , h ˙ , and a = u ¨ , v ¨ , r ¨ , h ¨ , the current state equation is inferred based on the state of the T − 1 frame in Equation (1). It shows the calculation of a new position and velocity for a given time interval, △ t , and acceleration, a .
p t = p t − 1 + Δ t v t − 1 + 0.5 a Δ t 2 v t = v t − 1 + a Δ t a t = a t − 1
Furthermore, the improved state transition matrix, A , after modification, is presented in Equation (2).
A = 1 Δ t 0.5 Δ t 2 0 1 Δ t 0 0 1
The state transition equation and observation equation of the updated uniformly accelerated Kalman filter model are shown in Equation (3).
x t = p v a t = 1 Δ t 0.5 Δ t 2 0 1 Δ t 0 0 1 p v a t − 1 + w t − 1 Z t = 1 0 0 p v a t − i T + V t
To sum up, we adopted the improved ByteTrack to realize the tracking of unauthorized UAVs. First, we categorized the detection results of PP-YOLO tiny into high-scoring boxes and low-scoring boxes based on the predefined high score threshold, Th, and the low score threshold, Tl. Then, we utilized the improved Kalman filtering to predict the current state of the target. The correlation between the detection results and the tracking trajectories was established by matching the detection results with the prediction results. In this study, we used the IOU to measure the similarity between the detection results and the prediction results and applied the Hungarian algorithm to solve the matching results. Subsequently, we matched the high-scoring boxes to the tracking trajectories and then matched the low-scoring boxes to the remaining unmatched trajectories. After these two rounds of matching, we established new tracking trajectories for the unmatched high-scoring boxes. For those unmatched tracking trajectories, we kept a buffer of 30 frames for re-matching.
Based on PP-YOLO tiny and the improved ByteTrack, we realized the detection and tracking of the unauthorized UAVs on the NVIDIA Jetson Xavier NX at 42.5 FPS, which satisfies the real-time processing requirements. It is important to emphasize that, in practical application contexts, we consider targets larger than 30 × 30 pixels as detectable. In our experiments, for non-cooperative UAVs with dimensions of 0.3 × 0.3 m, our system can achieve an effective detection distance of about 40 m under single-focal length conditions. However, it is crucial to note that the actual operational distance of the system against non-cooperative UAVs is primarily limited by the effective range of the net launching mechanism (3–8 m), which is significantly shorter than the system’s detection distance. Therefore, in practical operations, we focused more on the range of the net launching mechanism to ensure capture effectiveness.

4. Experiments and Results

In order to comprehensively assess the feasibility and applicability of the developed UAV Hunter system, we conducted hardware-in-the-loop simulation experiments for UAV detection and tracking and open-environment tests for UAV net capture.

4.1. The Hardware-in-the-Loop Simulation Experiments for UAV Detection and Tracking

Experimental platforms. We utilized a high-performance workstation equipped with four NVIDIA RTX 2080Ti GPUs to train the algorithm and then deployed it on the NVIDIA Jetson Xavier NX for inference. We adopted PaddlePaddle 2.3 for training, PaddleSlim 2.3 for quantization, and Paddle Inference 2.3 for inference [56]. The experimental platform is depicted in Figure 7, including an onboard computer and a remote control. The onboard computer performs UAV detection and tracking in the images and then transmits the processed results to the remote control in real time. We conducted hardware-in-the-loop simulation experiments for UAV detection and tracking on this platform to validate the system’s functionality.
Dataset. To train and test the algorithms, we prepared a dual-mode UAV detection and tracking dataset. For UAV detection, the dataset contained 3500 visible images and 2500 infrared images, of which 70% were used for training and 30% for testing. For UAV tracking, the dataset contained four sequences of visible images and three sequences of infrared images for a total of 2500 frames, all of which were used for testing. This dataset encompassed various UAV categories, including DJI Woo 1, DJI Royal 2, DJI Genie 3, and the self-built rotor UAV.

4.1.1. Results of UAV Detection Experiments

We initially trained PP-YOLO tiny until convergence on the workstation using the UAV detection training set and then deployed it on the onboard computer. Subsequently, the accuracy and real-time performance of the algorithm was evaluated on the airborne computer using the test set. Figure 8 illustrates the detection effects of the algorithm in visible mode for unauthorized UAVs. It was found that the algorithm accurately detects UAVs in complex environments with few false positives or false negatives, exhibiting exceptional adaptability and detection accuracy. Figure 9 shows the detection performance of the algorithms in infrared mode for unauthorized UAVs. Notably, the method achieves accurate detection of UAVs amidst various background interferences.
In addition, Table 2 displays the quantitative results of the detection experiment, emphasizing the performance in terms of accuracy and speed. The YOLO series of algorithms, specifically YOLOv4-tiny, YOLOv5-tiny, and YOLOv7-tiny, were assessed alongside a proposed algorithm denoted as “Ours”, which is based on the PP-YOLO-tiny architecture. The key metric for accuracy is Mean Average Precision (mAP), while the speed of the algorithms was measured in Frames Per Second (FPS). YOLOv4-tiny achieved an mAP of 92.6% with 27.3 FPS, demonstrating a reasonable balance between accuracy and speed. YOLOv5-tiny showed an improvement in accuracy, reaching an mAP of 94.4% while maintaining a good FPS of 32.8, indicating a slight compromise in speed for enhanced accuracy. YOLOv7-tiny further pushed the boundaries, achieving the highest accuracy with an mAP of 96.1% and an FPS of 34.3, illustrating its superior performance among the YOLO-tiny variants. However, the proposed algorithm outperformed all the other methods, achieving an mAP of 95.2% with 42.5 FPS on the test set. This result signifies that the PP-YOLO-tiny architecture not only maintains high detection accuracy but also exhibits significantly faster processing speeds compared to the other YOLO-tiny variants. This balance of high accuracy and speed renders the PP-YOLO-tiny algorithm an attractive choice for real-time object detection applications that require both precision and efficiency.
To more comprehensively assess the performance of our technique, particularly in complex environments, such as those with smoke or countermeasures, we designed a series of supplementary experiments. Specifically, we considered a variety of environmental factors that could affect the accuracy of visual inspection, with smoke and fog interference being particularly critical challenges. We simulated these complex scenes by adding smoke and cloud interference to visible images. We used Gaussian blur and noise to simulate the blur and graininess of smoke. The results are shown in Figure 10.
Cloud effects typically manifest as soft, diffuse white areas that may partially obscure image contents. To simulate this effect, we used Gaussian blur to create a smooth transition and increased the “whiteness” of the clouds by overlaying a semi-transparent white layer. The results are shown in Figure 11.
According to the simulation results, under the interference of smoke and clouds, we chose 100 images of each special scene as the test set. On this basis, we implemented a series of detection experiments, as shown in Table 3. These experiments covered not only standard test cases in clear visibility conditions, but also special scenarios for smoke and fog interference. By comparing and analyzing the results in these scenarios, we could more accurately assess the boundaries and potential performance of our technology.
After adding fog or smoke interference, the accuracy of the detection algorithm will decrease because such interference reduces the clarity and contrast of an image, causing the features of the target object to become blurred. This manifests as a decrease in the mAP value. Especially in cases of dense smoke or fog, the boundary and texture information of a target may be seriously lost, making it difficult for the detection algorithm to accurately identify the target. If the detection algorithm is fully optimized in hardware and the addition of interference does not significantly increase the complexity of image processing, the FPS value may remain unchanged or only fluctuate slightly.

4.1.2. Results of UAV Tracking Experiments

We adopted the improved ByteTrack to correlate the UAV detection results of the previous and the subsequent frames. The proposed method does not require training, so it can be directly applied to object detection results to achieve target tracking.
We validated the effectiveness of the proposed algorithm using multiple sequences of UAV flight images acquired in an open environment. Figure 12 and Figure 13, respectively, show the relevant tracking results in the visible and infrared modes, where the yellow curves indicate the motion trajectories of the tracked UAVs in the last 50 frames. It can be seen that the adopted algorithm can track the target accurately in both modes with different complex backgrounds and maintain target identity stability.
In the context of analyzing the performance of various tracking algorithms presented in Table 4, several noteworthy observations can be synthesized. The table comprehensively compares Sort, Deep Sort, OC Sort, ByteTrack, and the improved ByteTrack algorithm, using the metrics of Multiple Object Tracking Accuracy (MOTA), Multiple Object Tracking Precision based on the Intersection over the Union (MOTP-IoU), and Frames Per Second (FPS). The MOTAs of Deep Sort and OC Sort surpass that of Sort in terms of accuracy. Notably, the improved ByteTrack algorithm, achieved the highest MOTA of 92.1%, demonstrating a superior capability in accurately tracking UAVs over time. This underscores the significance of the proposed method in enhancing tracking effectiveness. The MOTP values indicate that all algorithms exhibit a high degree of precision in localizing tracked objects, with the proposed method once again leading the pack at 0.938. This further validates the improved method’s ability to precisely estimate the overlap between predicted and ground truth bounding boxes, which is crucial for accurate tracking. While Sort emerges as the fastest algorithm with an FPS of 31.0, Deep Sort and the improved ByteTrack achieve competitive speeds while maintaining significantly higher accuracy. Specifically, the improved ByteTrack, despite its improvements in accuracy, maintains an FPS of 29.1, highlighting its balance between performance and efficiency.

4.1.3. Results of the Image Return Experiments

The UAV detection and tracking results from the onboard computer are transmitted to the remote control and displayed on the attached screen in real time. Figure 14a and Figure 14b show the returned experimental results for real-time detection and tracking of UAVs in the visible and the infrared scenarios, respectively. With the help of these results, operators can respond more quickly and accurately to activate the tether-net capture device, deploying the rope net to apprehend unauthorized UAVs.

4.2. The Open-Environment Tests for UAV Net Capture

To test the functional integrity of the system, we conducted UAV net-capturing tests in the open environment. The test process is shown in Figure 15. Firstly, we used the ground station to plan the route and control the net-capturing UAV to fly to the target area. During flight, the net-capturing UAV utilized the onboard dual-mode sensors to collect real-time images, which were then processed by the onboard computer for UAV detection and tracking. The processing results were transmitted back to the ground remote control. The system alerted the operators when unauthorized UAVs were detected within the field of view. The operators manually adjusted the position and attitude of the net-capturing UAV with visual assistance until the target was placed in the center area of the image. At this point, the operators manually fired the airborne tether-net capture device to launch a rope net, completing the countermeasures against the unauthorized UAV.
The tests results for UAV net capture verify that the developed system is capable of effectively performing detection, tracking, and interception of unauthorized UAVs. Throughout the tests, the system exhibited stable operation, full functionality, and positive human–computer interaction, thus fulfilling the anticipated objectives.

5. Conclusions

To effectively deal with the harassment of unauthorized UAVs, we developed the UAV Hunter, a novel anti-UAV system based on a UAV platform. The system employs a six-rotor UAV as the base, equipped with a visual assist system and an airborne tether-net capture device for close-range interception of unauthorized UAVs. The system not only overcomes the spatial limitations of the traditional anti-UAV systems, but also enhances safety and efficiency. To improve the accuracy and success rate of net capture, we implemented real-time UAV detection and tracking on the NVIDIA Jetson Xavier NX using PPYOLO tiny and ByteTrack. The algorithm can work in both visible and infrared modes, with a frame rate of 42.5 FPS and an mAP of 95.2%, which provides precise visual localization information to the operators around the clock. To further enhance the human–machine interaction experience, we integrated the image display function and multiple control functions into a single remote control, which enables the operator to quickly obtain auxiliary positioning information and accurately launch the rope net remotely to intercept targets. The results of the hardware-in-the-loop simulations and open-environment tests further validate the functional completeness and practicality of the developed system. In the future, the developed system can be further expanded. For instance, by modifying the target recognition algorithm, the system can be utilized for both search-and-rescue operations in wildernesses, as well as for tracking down and apprehending fugitives.

Author Contributions

Conceptualization, R.L. and X.Y.; Data curation, T.Z., X.X. and B.T.; Formal analysis, X.Y.; Funding acquisition, R.L.; Investigation, X.X. and T.Z.; Methodology, X.X. and T.Z.; Resources, X.Y.; Software, T.Z. and J.F.; Supervision, R.L.; Validation, X.X. and J.F.; Visualization, T.Z. and B.T.; Writing—original draft, T.Z.; Writing—review and editing, X.X. and T.Z. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the Natural Science Foundation of China, under grants 62176263 and 62276274, and the Science Foundation for Distinguished Youth of Shaanxi Province, under grant 2021JC-35.

Data Availability Statement

The datasets used or analyzed during the current study are available from the corresponding author on reasonable request.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Xie, X.; Xi, J.; Yang, X.; Lu, R.; Xia, W. STFTrack: Spatio-Temporal-Focused Siamese Network for Infrared UAV Tracking. Drones 2023, 7, 296–318. [Google Scholar] [CrossRef]
  2. Fan, J.; Yang, X.; Lu, R.; Xie, X.; Li, W. Design and Implementation of Intelligent Inspection and Alarm Flight System for Epidemic Prevention. Drones 2021, 5, 68–82. [Google Scholar] [CrossRef]
  3. Zhao, M.; Xi, J.; Wang, L.; Xia, K.; Zheng, Y. Edge-Based Adaptive Secure Consensus for Nonlinear Multi-Agent Systems with Communication Link Attacks. Neurocomputing 2023, 551, 126505–126517. [Google Scholar] [CrossRef]
  4. Yang, Y.; Xiao, Y.; Li, T. Attacks on Formation Control for Multi-Agent Systems. IEEE Trans. Cybern. 2021, 52, 12805–12817. [Google Scholar] [CrossRef] [PubMed]
  5. Qian, B.; Yu, W.; Hao, H.; Zhu, H. Research on Development of Vehicular High Power Microwave Weapons. In Proceedings of the International Conference on Man-Machine-Environment System Engineering, Beijing, China, 21–23 October 2022; pp. 336–342. [Google Scholar]
  6. Yasmine, G.; Maha, G.; Hicham, M. Survey on Current Anti-Drone Systems: Process, Technologies, and Algorithms. Int. J. Syst. Syst. Eng. 2022, 12, 235–270. [Google Scholar] [CrossRef]
  7. Luo, H.; Lu, Y. Status and Development Trend of Foreign Anti-UAV of Low, Slow and Small Capability. Aerosp. Technol. 2019, 06, 32–36. [Google Scholar]
  8. Zhou, Z.; Cheng, C.; Zhang, X.; Bian, W. Numerical Simulation and Optimization on Launching Process of an Anti-UAV Capture. J. Ballist. 2019, 31, 1–6. [Google Scholar]
  9. Jiang, N.; Wang, K.; Peng, X.; Yu, X.; Wang, Q.; Xing, J.; Li, G.; Guo, G.; Ye, Q.; Jiao, J.; et al. Anti-UAV: A Large-Scale Benchmark for Vision-Based UAV Tracking. IEEE Trans. Multimed. 2021, 25, 486–500. [Google Scholar] [CrossRef]
  10. Bo, C.; Wei, Y.; Wang, X.; Shi, Z.; Xiao, Y. Vision-Based Anti-UAV Detection Based on YOLOv7-GS in Complex Backgrounds. Drones 2024, 8, 331. [Google Scholar] [CrossRef]
  11. Zhao, J.; Zhang, J.; Li, D.; Wang, D. Vision-Based Anti-UAV Detection and Tracking. IEEE Trans. Intell. Transp. Syst. 2022, 23, 25323–25334. [Google Scholar] [CrossRef]
  12. Hu, Y.; Wu, X.; Zheng, G.; Liu, X. Object Detection of UAV for Anti-UAV Based on Improved YOLO v3. In Proceedings of the 2019 Chinese Control Conference (CCC), Guangzhou, China, 27–30 July 2019; pp. 8386–8390. [Google Scholar]
  13. Zhang, J.; Lin, Y.; Zhou, X.; Shi, P.; Zhu, X.; Zeng, D. Precision in Pursuit: A Multi-consistency Joint Approach for Infrared Anti-UAV Tracking. Vis. Comput. 2024. [Google Scholar] [CrossRef]
  14. Huang, M.; Mi, W.; Wang, Y. EDGS-YOLOv8: An Improved YOLOv8 Lightweight UAV Detection Model. Drones 2024, 8, 337. [Google Scholar] [CrossRef]
  15. Fang, H.; Liao, Z.; Wang, X.; Chang, Y.; Yan, L. Differentiated Attention Guided Network over Hierarchical and Aggregated Features for Intelligent UAV Surveillance. IEEE Trans. Ind. Inform. 2023, 19, 9909–9920. [Google Scholar] [CrossRef]
  16. Zhang, T.; Yang, X.; Lu, R.; Xie, X.; Wang, S.; Su, S. Context-Aware DGCN-Based Ship Formation Recognition in Remote Sensing Images. Remote Sens. 2024, 16, 3435. [Google Scholar] [CrossRef]
  17. Xiang, H.; Liang, C.; Qiao, Z.; Yuan, X.; Cao, G. Parameters Simulation and Optimization of Flying Net for UAVs Interception. IEEE Access 2022, 10, 56668–56676. [Google Scholar] [CrossRef]
  18. Wei, Y.; Du, Z.; Chen, X.; Ma, R.; Shi, W. Dynamic Simulation Analysis of Net-capturing UAV Based on ABAQUS. J. Ordnance Equip. Eng. 2021, 42, 98–103. [Google Scholar]
  19. Wang, L.; Mi, S.; Song, Q. A Portable Anti-UAV Intelligent Interception Network System. China, CN105501436A, 20 April 2016. [Google Scholar]
  20. Su, G.; Xiao, Z.; Li, R. Shield Portable Anti-UAV Missile Soft Damage Recovery Device. China, CN106767172A, 31 May 2017. [Google Scholar]
  21. Garcia, A.; Lee, J.; Kim, D. Anti-drone System: A Visual-based Drone Detection using Neural Networks. In Proceedings of the International Conference on Information and Communication Technology Convergence, Jeju, Republic of Korea, 21–23 October 2020; pp. 559–561. [Google Scholar]
  22. Fang, G.; Yi, J.; Wan, X.; Liu, Y.; Ke, H. Experimental Research of Multi-static Passive Radar with a Single Antenna for Drone Detection. IEEE Access 2018, 6, 33542–33551. [Google Scholar] [CrossRef]
  23. Sabogal, A.; Gómez, M.; Ugarte, J. Heterogeneous Acoustic Features Space for Automatic Classification of Drone Audio Signals. In Proceedings of the 8th Workshop on Engineering Applications, Medellín, Colombia, 6–8 October 2021; pp. 97–109. [Google Scholar]
  24. Bosio, I.; Garibotto, C.; Lavagetto, F.; Scriarrone, A.; Zappatore, S. Unauthorized Amateur UAV Detection Based on WiFi Statistical Fingerprint Analysis. IEEE Commun. Mag. 2018, 56, 106–111. [Google Scholar] [CrossRef]
  25. Redmon, J.; Divvala, S.; Girshick, R.; Farhadi, A. You Only Look Once: Unified, Real-time Object Detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Seattle, WA, USA, 27–30 June 2016; pp. 779–788. [Google Scholar]
  26. Redmon, J.; Farhadi, A. YOLO9000: Better, Faster, Stronger. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA, 21–26 July 2017; pp. 6517–6525. [Google Scholar]
  27. Redmon, J.; Farhadi, A. YOLOv3: An Incremental Improvement. arXiv 2018, arXiv:1804.02767. [Google Scholar]
  28. Bochkovskiy, A.; Wang, C.; Liao, H. YOLO V4: Optimal Speed and Accuracy of Object Detection. arXiv 2020, arXiv:2004.10934. [Google Scholar]
  29. Liu, W.; Anguelov, D.; Erhan, D.; Szegedy, C.; Reed, S.; Fu, C.; Berg, A.C. SSD: Single Shot Multibox Detector. In Proceedings of the 14th European Conference on Computer Vision, Amsterdam, The Netherlands, 8–16 October 2016; pp. 21–37. [Google Scholar]
  30. Fu, C.; Liu, W.; Ranga, A.; Berg, A. DSSD: Deconvolutional Single Shot Detector. arXiv 2017, arXiv:1701.06659. [Google Scholar]
  31. Duan, K.; Bai, S.; Xie, L.; Qi, H.; Huang, Q.; Tian, Q. CenterNet: Keypoint Triplets for Object Detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, Seoul, Republic of Korea, 27 October−2 November 2019; pp. 6569–6578. [Google Scholar]
  32. Tang, J. Research on Vision-Based Detection and Tracking Technology for Small UAV. Master’s Thesis, Nanjing University of Science & Technology, Nanjing, China, 2022. [Google Scholar]
  33. Shi, Q. Design of UAV Intrusion Detection System Based on Machine Vision. Master’s Thesis, Guilin University of Electronic Technology, Guilin, China, 2021. [Google Scholar]
  34. Jing, X. A Master Thesis Submitted to University of Electronic Science and Technology of China. Master’s Thesis, University of Electronic Science and Technology of China, Chengdu, China, 2021. [Google Scholar]
  35. Zhang, R.; Li, N.; Zhang, X. Low-altitude UAV Detection Method Based on Optimized CenterNet. J. Beijing Univ. Aeronaut. Astronaut. 2021, 48, 2335–2344. [Google Scholar]
  36. Liu, P. Research on Low-Altitude UAV Target Detection Based on Improved SSD Detection Network. Master’s Thesis, University of Chinese Academy of Sciences, Beijing, China, 2020. [Google Scholar]
  37. Zhou, Q.; Xia, M. Research on Multi-UAV Detection Based on Improved SSD Algorithm. Inf. Technol. 2020, 4, 71–76. [Google Scholar]
  38. Ding, M.; Li, N.; Song, Z.; Zhang, R.; Zhang, X.; Zhou, H. A Lightweight Action Recognition Method for Unmanned-Aerial-Vehicle Video. In Proceedings of the IEEE 3rd International Conference on Electronics and Communication Engineering, Piscataway, NJ, USA, 14–16 December 2020; pp. 181–185. [Google Scholar]
  39. Long, X.; Deng, K.; Wang, G.; Zhang, Y.; Dang, Q.; Gao, Y. PP-YOLO: An Effective and Efficient Implementation of Object Detector. arXiv 2020, arXiv:2007.12099. [Google Scholar]
  40. Pan, S.; Tong, Z.; Zhao, Y.; Zhao, Z.; Su, F.; Zhuang, B. Multi-Object Tracking Hierarchically in Visual Data Taken from Drones. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, Seoul, Republic of Korea, 27 October–2 November 2019; pp. 135–143. [Google Scholar]
  41. Khalkhali, M.; Vahedian, A.; Yazdi, S. Situation Assessment-Augmented Interactive Kalman filter for Multi-Vehicle Tracking. IEEE Trans. Intell. Transp. Syst. 2021, 23, 3766–3776. [Google Scholar] [CrossRef]
  42. Li, Y.; Fu, C.; Huang, Z.; Zhang, Y.; Pan, J. Intermittent Contextual Learning for Key Filter-Aware UAV Object Tracking using Deep Convolutional Feature. IEEE Trans. Multimed. 2020, 23, 810–822. [Google Scholar] [CrossRef]
  43. Li, S.; Yang, X.; Wang, X.; Zeng, D.; Ye, H.; Zhao, Q. Learning Target-Aware Vision Transformers for Real-Time UAV Tracking. IEEE Trans. Geosci. Remote Sens. 2024, 62, 1–18. [Google Scholar] [CrossRef]
  44. Fang, H.; Wu, C.; Wang, X.; Zhou, F.; Chang, Y.; Yan, L. Online Infrared UAV Target Tracking with Enhanced Context-Awareness and Pixel-Wise Attention Modulation. IEEE Trans. Geosci. Remote Sens. 2024, 62, 1–17. [Google Scholar] [CrossRef]
  45. Sun, N.; Zhao, J.; Shi, Q.; Liu, C.; Liu, P. Moving Target Tracking by Unmanned Aerial Vehicle: A Survey and Taxonomy. IEEE Trans. Ind. Inform. 2024, 20, 7056–7068. [Google Scholar] [CrossRef]
  46. Gao, Y.; Gan, Z.; Chen, M.; Ma, H.; Mao, X. Hybrid Dual-Scale Neural Network Model for Tracking Complex Maneuvering UAVs. Drones 2024, 8, 3. [Google Scholar] [CrossRef]
  47. Wu, X.; Li, W.; Hong, D.; Ran, T.; Du, Q. Deep learning for Unmanned Aerial Vehicle-based Object Detection and Tracking: A Survey. IEEE Geosci. Remote Sens. Mag. 2021, 10, 91–124. [Google Scholar] [CrossRef]
  48. Osco, L.; Junior, J.; Ramos, A.; Jorge, L.; Fatholahi, S.; Silva, J.; Matsubara, E.; Pistori, H.; Goncalves, W.; Li, J. A Review on Deep Learning in UAV Remote Sensing. Int. J. Appl. Earth Obs. Geo-Inf. 2021, 102, 102456. [Google Scholar] [CrossRef]
  49. Srivastava, S.; Narayan, S.; Mittal, S. A Survey of Deep Learning Techniques for Vehicle Detection from UAV Images. J. Syst. Archit. 2021, 117, 102152. [Google Scholar] [CrossRef]
  50. Cazzato, D.; Cimarelli, C.; Sanchez, L.; Jose, L.; Voos, H.; Leo, M. A Survey of Computer Vision Methods for 2D Object Detection from Unmanned Aerial Vehicle. J. Imaging 2020, 6, 78–116. [Google Scholar] [CrossRef] [PubMed]
  51. Li, F.; Fu, C.; Lin, F.; Li, Y.; Lu, P. Training-set Distillation for Real time UAV Object Tracking. In Proceedings of the IEEE International Conference on Robotics and Automation, Paris, France, 31 May–16 June 2020; pp. 9715–9721. [Google Scholar]
  52. Zhang, Y.; Sun, P.; Jiang, Y.; Yu, D.; Weng, F.; Yuan, Z.; Luo, P.; Liu, P.; Liu, W.; Wang, X. ByteTrack: Multi-Object Tracking by Associating Every Detection Box. In Proceedings of the European Conference on Computer Vision, Tel Aviv, Israel, 24–28 October 2022; pp. 1–21. [Google Scholar]
  53. He, K.; Zhang, X.; Ren, S.; Sun, J. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition. IEEE Trans. Pattern Anal. Mach. Intell. 2015, 37, 1904–1916. [Google Scholar] [CrossRef] [PubMed]
  54. Yu, J.; Jiang, Y.; Wang, Z.; Cao, Z.; Huang, T. Unitbox: An Advanced Object Detection Network. In Proceedings of the 24th ACM International Conference on Multimedia, Amsterdam, The Netherlands, 15–19 October 2016; pp. 516–520. [Google Scholar]
  55. Howard, A.; Sandler, M.; Chu, G.; Chen, L.; Chen, B.; Tan, M. Searching for Mobilenetv3. In Proceedings of the IEEE International Conference on Computer Vision, Seoul, Republic of Korea, 27 October–2 November 2019; pp. 1314–1324. [Google Scholar]
  56. Ma, Y.; Yu, D.; Wu, T.; Wang, H. PaddlePaddle: An Open-Source Deep Learning Platform from Industrial Practice. Front. Data Comput. 2019, 1, 105–115. [Google Scholar]
Drones 08 00573 g001
Figure 1. Framework of the UAV Hunter system.
Figure 1. Framework of the UAV Hunter system.
Drones 08 00573 g001
Drones 08 00573 g002
Figure 2. Hardware framework of the net-capturing UAV.
Figure 2. Hardware framework of the net-capturing UAV.
Drones 08 00573 g002
Drones 08 00573 g003
Figure 3. Hardware configuration of our assembled six-rotor drone.
Figure 3. Hardware configuration of our assembled six-rotor drone.
Drones 08 00573 g003
Drones 08 00573 g004
Figure 4. Overall configuration of the airborne tether-net capture device.
Figure 4. Overall configuration of the airborne tether-net capture device.
Drones 08 00573 g004
Drones 08 00573 g005
Figure 5. Framework of the UAV detection and tracking algorithm.
Figure 5. Framework of the UAV detection and tracking algorithm.
Drones 08 00573 g005
Drones 08 00573 g006
Figure 6. The main framework of PP-YOLO tiny.
Figure 6. The main framework of PP-YOLO tiny.
Drones 08 00573 g006
Drones 08 00573 g007
Figure 7. The hardware structure for the simulation experiment platform. (a–c) are the onboard computer, remote controller, and dual-mode sensor pod, respectively.
Figure 7. The hardware structure for the simulation experiment platform. (a–c) are the onboard computer, remote controller, and dual-mode sensor pod, respectively.
Drones 08 00573 g007
Drones 08 00573 g008aDrones 08 00573 g008b
Figure 8. Detection results for the UAVs in the visible mode. The red boxes represent the detected non-cooperating UAVs.
Figure 8. Detection results for the UAVs in the visible mode. The red boxes represent the detected non-cooperating UAVs.
Drones 08 00573 g008aDrones 08 00573 g008b
Drones 08 00573 g009
Figure 9. Detection results for the UAVs in the infrared mode. The red boxes represent the detected non-cooperating UAVs.
Figure 9. Detection results for the UAVs in the infrared mode. The red boxes represent the detected non-cooperating UAVs.
Drones 08 00573 g009
Drones 08 00573 g010
Figure 10. The simulation test set for smoke jamming effects. (a) The ground truth. We used Gaussian blur to simulate the blur effect of smoke and added Gaussian noise to simulate particles of smoke. The intensities of the smoke effects in (b–e) are 0.01, 0.01, 0.01, and 0.02, respectively, and the corresponding noise intensities are 0.001, 0.002, 0.003, and 0.002.
Figure 10. The simulation test set for smoke jamming effects. (a) The ground truth. We used Gaussian blur to simulate the blur effect of smoke and added Gaussian noise to simulate particles of smoke. The intensities of the smoke effects in (b–e) are 0.01, 0.01, 0.01, and 0.02, respectively, and the corresponding noise intensities are 0.001, 0.002, 0.003, and 0.002.
Drones 08 00573 g010
Drones 08 00573 g011
Figure 11. The simulation test set for cloud jamming effects. (a) The ground truth. We used Gaussian blur to create a smooth transition and increased the “whiteness” of the clouds by overlaying a semi-transparent white layer. The intensities of the cloud effects in (b–e) are 0.7, 0.8, 0.9, and 0.9, respectively, intensity mainly controlling the transparency of a cloud, and the extents of cloud coverage are 0.3, 0.4, 0.5, and 0.7, extent mainly controlling the proportion of cloud in an image.
Figure 11. The simulation test set for cloud jamming effects. (a) The ground truth. We used Gaussian blur to create a smooth transition and increased the “whiteness” of the clouds by overlaying a semi-transparent white layer. The intensities of the cloud effects in (b–e) are 0.7, 0.8, 0.9, and 0.9, respectively, intensity mainly controlling the transparency of a cloud, and the extents of cloud coverage are 0.3, 0.4, 0.5, and 0.7, extent mainly controlling the proportion of cloud in an image.
Drones 08 00573 g011
Drones 08 00573 g012
Figure 12. Tracking results for the UAVs in the visible mode.
Figure 12. Tracking results for the UAVs in the visible mode.
Drones 08 00573 g012
Drones 08 00573 g013
Figure 13. Tracking results for the UAVs in the infrared mode.
Figure 13. Tracking results for the UAVs in the infrared mode.
Drones 08 00573 g013
Drones 08 00573 g014
Figure 14. Return results displayed on the remote control. The red boxes represent the detected UAV target, and the yellow line refers to the tracking trajectory.
Figure 14. Return results displayed on the remote control. The red boxes represent the detected UAV target, and the yellow line refers to the tracking trajectory.
Drones 08 00573 g014
Drones 08 00573 g015
Figure 15. Test process of UAV net capture. (a) Planning the path and controlling the UAV’s flight. (b) Detecting and tracking an unauthorized UAV. (c) Launching the rope net and capturing the UAV. (d) Recovering the net.
Figure 15. Test process of UAV net capture. (a) Planning the path and controlling the UAV’s flight. (b) Detecting and tracking an unauthorized UAV. (c) Launching the rope net and capturing the UAV. (d) Recovering the net.
Drones 08 00573 g015
Table 1. The component list of the net-capturing UAV hardware.
Table 1. The component list of the net-capturing UAV hardware.
SystemsComponent ListSpecification
Net-capturing UAVFlight controllerCUAV-V5+
Electronic speed controlXRotor-40A
BLDC motorsX4112S
Digital transmission module3DR·V5-Radio
GNSS moduleCUAV-NEO v2
BatteriesTATTU-10000mAh-25C-22.2V-6S1P
Onboard computerNVIDIA Jetson Xavier NX
Dual-mode sensor podSHD10T3 Visible-IR Pod
Remote controllerSIYI-MK15
Table 2. The quantitative experimental results for detection.
Table 2. The quantitative experimental results for detection.
Parameter/MethodYOLOv4 TinyYOLOv5 TinyYOLOv7 TinyOurs 1
mAP (%) ↑ 92.694.496.195.2
FPS ↑ 27.332.834.342.5
1 Ours refer to the PP-YOLO tiny algorithm.
Table 3. The quantitative experimental results for detection. “Light Smoke Interference” refers to an intensity of smoke effect and a noise intensity less than 0.01 and 0.001, respectively, and higher values are classed as “Heavy Smoke Interference”. “Light Fog Interference” means that the transparency of clouds and the coverage area of clouds are less than 0.7 and 0.4, respectively. Other conditions with higher values are classed as “Heavy Fog Interference”.
Table 3. The quantitative experimental results for detection. “Light Smoke Interference” refers to an intensity of smoke effect and a noise intensity less than 0.01 and 0.001, respectively, and higher values are classed as “Heavy Smoke Interference”. “Light Fog Interference” means that the transparency of clouds and the coverage area of clouds are less than 0.7 and 0.4, respectively. Other conditions with higher values are classed as “Heavy Fog Interference”.
MethodYOLOv4 TinyYOLOv5 TinyYOLOv7 TinyOurs 1
mAP ↑ FPS ↑ mAP ↑ FPS ↑ mAP ↑ FPS ↑ mAP ↑ FPS ↑
Clear Visibility92.627.394.432.896.134.395.242.5
Light Smoke Interference88.427.390.130.993.234.391.342.5
Heavy Smoke Interference81.227.182.929.988.433.684.741.8
Light Fog Interference87.327.489.732.692.734.090.142.3
Heavy Fog Interference80.726.982.631.189.932.887.541.6
1 Ours refer to the PP-YOLO tiny algorithm.
Table 4. The quantitative experimental results for tracking.
Table 4. The quantitative experimental results for tracking.
Parameter/MethodSortDeep SortOC-SortByteTrackOurs 1
MOTA % ↑ 85.590.891.590.392.1
MOTP (IoU) ↑ 0.8220.9250.9370.9320.938
FPS ↑ 31.021.627.429.329.1
1 Ours refer to the improved ByteTrack algorithm.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Zhang, T.; Lu, R.; Yang, X.; Xie, X.; Fan, J.; Tang, B. UAV Hunter: A Net-Capturing UAV System with Improved Detection and Tracking Methods for Anti-UAV Defense. Drones 2024, 8, 573. https://doi.org/10.3390/drones8100573

AMA Style

Zhang T, Lu R, Yang X, Xie X, Fan J, Tang B. UAV Hunter: A Net-Capturing UAV System with Improved Detection and Tracking Methods for Anti-UAV Defense. Drones. 2024; 8(10):573. https://doi.org/10.3390/drones8100573

Chicago/Turabian Style

Zhang, Tao, Ruitao Lu, Xiaogang Yang, Xueli Xie, Jiwei Fan, and Bin Tang. 2024. "UAV Hunter: A Net-Capturing UAV System with Improved Detection and Tracking Methods for Anti-UAV Defense" Drones 8, no. 10: 573. https://doi.org/10.3390/drones8100573

APA Style

Zhang, T., Lu, R., Yang, X., Xie, X., Fan, J., & Tang, B. (2024). UAV Hunter: A Net-Capturing UAV System with Improved Detection and Tracking Methods for Anti-UAV Defense. Drones, 8(10), 573. https://doi.org/10.3390/drones8100573

Article Metrics

Back to TopTop