In recent years, the rapid advancement of unmanned aerial vehicle technology has attracted widespread attention. Unmanned aerial vehicles have been widely applied in military and civilian domains, ranging from logistics delivery and disaster relief to urban traffic management and military reconnaissance. Unmanned aerial vehicle formations can collaboratively cover larger geographic areas and obtain more comprehensive reconnaissance information, becoming an important development trend. To accomplish cooperative reconnaissance tasks, efficient trajectory planning and precise video data transmission are essential. In my research, I focus on addressing the challenges in trajectory planning and object detection for unmanned aerial vehicle formations.
The operation of unmanned aerial vehicle formations in complex three-dimensional urban or battlefield environments requires the integration of advanced trajectory planning and object detection capabilities. Traditional methods, such as artificial potential field methods and swarm intelligence algorithms, often suffer from local optima, poor adaptability to dynamic obstacles, and insufficient coordination in multi-unmanned aerial vehicle systems. Deep reinforcement learning offers a promising alternative, but existing algorithms frequently encounter slow convergence and oscillatory trajectories. Moreover, the limited wireless bandwidth in video transmission poses a critical bottleneck for real-time monitoring. To address these issues, I propose novel methods for trajectory planning and region-of-interest selection, and I design a practical edge-computing software for object detection. In this article, I present my research methodology, key algorithms, experimental results, and system implementation.
1. Research Background and Significance
Unmanned aerial vehicle formations are increasingly deployed in various fields, including military reconnaissance, traffic surveillance, logistics delivery, and disaster response. In a typical low-altitude mission, a group of unmanned aerial vehicles must navigate through dense urban canyons, avoid both static and moving obstacles, and arrive at designated areas while maintaining a safe formation. Meanwhile, the onboard camera system needs to detect and track objects of interest, such as vehicles or people, and transmit only the most relevant video data to the ground control station to conserve bandwidth. These two tasks—trajectory planning and object detection—are deeply coupled in real-world scenarios. A well-planned path enables robust visual observation, and efficient object detection helps the unmanned aerial vehicle focus on critical regions, thereby reducing unnecessary data transmission.
The significance of this research is twofold. First, by integrating trajectory planning with object detection, I aim to improve the safety and efficiency of unmanned aerial vehicle formations in low-altitude airspace. Second, the developed algorithms and software provide a practical technical foundation for real-world applications such as smart city patrols, border security, and traffic management. The research addresses the conflict between wide-area coverage and high-resolution transmission, and it shifts the paradigm from passive monitoring to active intelligent alerting.
2. Literature Review
2.1 Deep Reinforcement Learning for Trajectory Planning
Deep reinforcement learning has emerged as a powerful tool for path planning in complex environments. Prior works have applied algorithms such as Soft Actor-Critic (SAC), Deep Deterministic Policy Gradient (DDPG), Twin Delayed DDPG (TD3), and Proximal Policy Optimization (PPO) to unmanned aerial vehicle navigation. Table 1 summarizes several representative approaches. These methods generally formulate the path planning problem as a Markov Decision Process and optimize a cumulative reward function.
| Type | DRL Algorithm | Reference | Advantages | Challenges |
|---|---|---|---|---|
| Single-UAV | SAC with Hindsight Experience Replay | [1] | Sample efficiency, exploration | High computational cost |
| Single-UAV | Action Decoupling SAC | [2] | Handles hybrid action spaces | Convergence speed |
| Single-UAV | DDQN with random obstacle training | [3] | Robust to random obstacles | Discrete actions only |
| Multi-UAV | MADDPG | [4] | Cooperative learning | Large communication overhead |
| Multi-UAV | MATD3 with task decomposition | [5] | Better exploration in complex environments | Long training time |
Although these methods have achieved promising results, they often neglect the collision safety among unmanned aerial vehicles in a formation and produce trajectories with unnecessary oscillations. My proposed SPER-TD3 algorithm aims to overcome these limitations by enhancing training stability and trajectory smoothness.
2.2 Object Detection Algorithms
Object detection has undergone a major paradigm shift from hand-crafted feature methods to deep learning-based approaches. Two-stage detectors such as Faster R-CNN and Cascade R-CNN achieve high accuracy but at the cost of inference speed. Single-stage detectors, particularly the YOLO family, have become the de facto standard for real-time unmanned aerial vehicle vision tasks due to their excellent balance between speed and accuracy. Table 2 compares several representative detectors.
| Detector | Backbone | mAP | FPS | Remarks |
|---|---|---|---|---|
| Faster R-CNN | ResNet-50 | 78 | 7 | Two-stage |
| SSD | VGG-16 | 77 | 22 | Multi-scale |
| YOLOv4 | CSPDarknet53 | 82 | 45 | Efficient |
| YOLOv7 | CSPDarknet | 85 | 35 | Bag-of-freebies |
| YOLOv10 | C2f-based | 86 | 50 | NMS-free |
YOLOv10, the latest version, introduces a dual assignment strategy that eliminates the need for non-maximum suppression, significantly reducing inference latency. This makes it particularly suitable for deployment on edge computing platforms such as the Jetson AGX Orin, which I use in my system.
3. Related Theoretical Foundations
3.1 Convolutional Neural Networks
Convolutional neural networks are the building blocks of modern object detection. In a CNN, the convolution operation extracts local features using learnable filters. For a two-dimensional input image \(I\) and a convolution kernel \(K\), the output feature map can be expressed as:
$$(I * K)(x,y) = \sum_m \sum_n I(x+m, y+n) K(m,n)$$
Pooling layers reduce the spatial dimensionality while preserving important information. Max pooling is defined as:
$$O(i,j) = \max_{(p,q) \in R} I(p,q)$$
where \(R\) is the pooling window. Average pooling is given by:
$$O(i,j) = \frac{1}{k} \sum_{(p,q) \in R} I(p,q)$$
The fully connected layer performs a linear transformation followed by an activation function, producing class scores or regression values.
3.2 Reinforcement Learning and Markov Decision Processes
Reinforcement learning is formalized as a Markov Decision Process (MDP) defined by the tuple \(\langle S, A, P, R, \gamma \rangle\). An agent observes state \(s_t\), selects action \(a_t\), receives reward \(r_t\), and transitions to state \(s_{t+1}\). The goal is to maximize the expected discounted cumulative reward. The temporal-difference update for a value function \(V(s_t)\) is:
$$V(s_t) \leftarrow V(s_t) + \alpha \left[ r_{t+1} + \gamma V(s_{t+1}) – V(s_t) \right]$$
Q-learning uses an action-value function \(Q(s,a)\) and updates it according to:
$$Q(s_t,a_t) \leftarrow Q(s_t,a_t) + \alpha \left[ r_{t+1} + \gamma \max_a Q(s_{t+1}, a) – Q(s_t,a_t) \right]$$
Deep Q-Networks approximate the Q-function with a deep neural network and employ experience replay and target networks to stabilize training.
3.3 UAV Formation Control Methods
Several formation control methods exist, including leader-follower, virtual leader, and behavior-based methods. Table 3 highlights their characteristics.
| Method | Advantages | Disadvantages | Applicability |
|---|---|---|---|
| Leader-Follower | Simple, easy to implement | Single point of failure | Small-scale formations |
| Virtual Leader | Strong formation rigidity | High communication burden | Static environments |
| Behavior-based | High flexibility and adaptability | Coordination can be weak | Dynamic environments |
In my work, I design an angle-based virtual leader method that allows flexible adjustment of the formation shape while maintaining robustness against individual follower failures.
4. Proposed SPER-TD3 for Formation Trajectory Planning
4.1 Problem Formulation
I consider a scenario where a formation of unmanned aerial vehicles operates in a three-dimensional environment with static and dynamic obstacles. The environment model includes spheres, cylinders, and cones. The collision condition is expressed by the obstacle function \(\Phi\):
$$
\Phi =
\begin{cases}
\left(\frac{x-x_{sp}}{r}\right)^2 + \left(\frac{y-y_{sp}}{r}\right)^2 + \left(\frac{z-z_{sp}}{r}\right)^2 < 1 & \text{(sphere)} \\
\left(\frac{x-x_{cy}}{r}\right)^2 + \left(\frac{y-y_{cy}}{r}\right)^2 < 1, \; 0 \le z \le H & \text{(cylinder)} \\
\left(\frac{x-x_{co}}{r}\right)^2 + \left(\frac{y-y_{co}}{r}\right)^2 < \left(1 – \frac{z}{H}\right)^2 & \text{(cone)}
\end{cases}
$$
The motion model of a rotorcraft unmanned aerial vehicle is described by the heading angle \(\psi\) and flight path angle \(\gamma\):
$$
\psi = \arcsin\left(\frac{\Delta y}{\sqrt{\Delta x^2 + \Delta y^2}}\right), \quad
\gamma = \arcsin\left(\frac{\Delta z}{\sqrt{\Delta x^2 + \Delta y^2 + \Delta z^2}}\right)
$$
The position update from time \(t\) to \(t+1\) is given by:
$$
\begin{aligned}
\Delta x_{res} &= R \cos(\psi_{res}) \cos(\gamma_{res}) \\
\Delta y_{res} &= R \cos(\psi_{res}) \sin(\gamma_{res}) \\
\Delta z_{res} &= R \sin(\gamma_{res})
\end{aligned}
$$
where \(R\) is the step size, and the subscript \(res\) indicates kinematic constraints.
4.2 Formation Model with Angle-Based Virtual Leader
To maintain the formation, I compute the direction vector \(\mathbf{F}\) of the leader:
$$\mathbf{F} = \frac{q_{t+1} – q_t}{\|q_{t+1} – q_t\|}$$
Then a rotation axis \(\mathbf{V} = \mathbf{F} \times \mathbf{K}\) is calculated, where \(\mathbf{K}\) is a non-parallel unit vector. After normalizing \(\mathbf{V}\) to \(\mathbf{V}’\), the Rodrigues rotation formula gives the rotated vector:
$$\mathbf{V}_{rot} = \mathbf{V} \cos\theta + (\mathbf{V}’ \times \mathbf{V}) \sin\theta + \mathbf{V}'(\mathbf{V}’ \cdot \mathbf{V})(1-\cos\theta)$$
The virtual leader position is then obtained as:
$$Q = q_t + D \mathbf{V}_{rot}$$
where \(D\) is the desired distance between the leader and the virtual leader. Each follower follows its own virtual leader, which improves fault tolerance.
To smooth the follower trajectories, I introduce a curvature-based dynamic smoothing algorithm. The instantaneous curvature is:
$$\theta_{curvature} = \arccos\left(\frac{\mathbf{F}_{ave} \cdot \mathbf{F}’}{\|\mathbf{F}_{ave}\| \|\mathbf{F}’\|}\right)$$
The smoothing factor \(\alpha\) is adjusted as:
$$
\alpha = \begin{cases}
0.5 & \text{if } \theta_{curvature} > \frac{\pi}{4} \\
1 & \text{otherwise}
\end{cases}
$$
Then the smoothed direction is:
$$\mathbf{F}_{smooth} = \alpha \mathbf{F}_{ave} + (1-\alpha) \mathbf{F}’$$
and the follower position is updated by \(q_{t+1} = q_t + v \mathbf{F}_{smooth}\).
4.3 Optimization Problem
The trajectory planning objective is to minimize a weighted cost function that accounts for path length, collision probability, and formation keeping:
$$\min_{q_1, q_2, \dots, q_N} \; \omega_1 \sum_{i=1}^{N-1} \|q_{i+1} – q_i\| + \omega_2 P_{collision} + \omega_3 \sum_{i=1}^{N} \|q_{i} – q_{i,leader}\|$$
subject to kinematics constraints and obstacle avoidance constraints:
$$
\begin{aligned}
|\psi| \le \psi_{max}, \quad |\gamma| \le \gamma_{max}, \\
q_i \in \mathcal{A}, \quad \|q_i – o\| \ge d_{min}, \quad \forall o \in \mathcal{O}
\end{aligned}
$$
where \(\mathcal{A}\) is the safe airspace and \(\mathcal{O}\) is the obstacle set.
4.4 Markov Decision Process Modeling
I model the problem as an MDP with:
- State space \(S\) – For each obstacle \(i\), the state is
$$s_i = [\Delta x_i, \Delta y_i, \Delta z_i, \Delta x_{goal}, \Delta y_{goal}, \Delta z_{goal}]$$
- Action space \(A\) – The action is the steering angle \(\eta_i\) generated by the TD3 network for each obstacle.
- Reward function – I design a composite reward:
$$R_{total} = R_{col} + R_{len} + R_{ang}$$
The collision reward is:
$$
R_{col} = \begin{cases}
-1 & \text{if collision} \\
R_s – k d(q_t, o) & \text{if } d(q_t, o) < d_{safe} \\
0 & \text{otherwise}
\end{cases}
$$
The distance-to-goal reward is:
$$
R_{len} = \begin{cases}
3 & \text{if } d(q_t, q_{goal}) < d_{threshold} \\
-\frac{d(q_t, q_{goal})}{d(q_{start}, q_{goal})} & \text{otherwise}
\end{cases}
$$
The angle smoothing reward penalizes large changes in heading and flight path angles:
$$R_{ang} = -0.1 \left( \frac{\Delta \psi}{\psi_{max}} + \frac{\Delta \gamma}{\gamma_{max}} \right)$$
4.5 SPER-TD3 Algorithm
I propose a SumTree-based prioritized experience replay integrated with the TD3 algorithm. The key idea is to assign higher sampling probabilities to transitions with larger TD errors, enabling the unmanned aerial vehicle to learn more efficiently from important experiences. The SumTree structure stores priorities at leaf nodes and internal nodes store cumulative sums, allowing efficient sampling. The overall framework is illustrated conceptually in the pseudo-code below.
| Algorithm: SPER-TD3 |
|---|
|
1. Initialize Actor network \(\theta\), Critic networks \(\phi_1, \phi_2\), target networks \(\theta’ \leftarrow \theta\), \(\phi’_1 \leftarrow \phi_1\), \(\phi’_2 \leftarrow \phi_2\). 2. Initialize SumTree priority buffer with priority exponent \(\alpha_p\) and importance sampling exponent \(\beta\). 3. Initialize hyperparameters: \(\gamma\), \(\tau\), learning rates, noise parameters, policy delay. 4. Repeat for each episode: 5. Observe state \(s\). 6. Select action \(a = \pi(s;\theta) + \epsilon\), where \(\epsilon \sim \mathcal{N}(0,\sigma^2)\). 7. Execute \(a\), observe reward \(r\) and next state \(s’\). 8. Compute TD error: \(\delta = r + \gamma \min_j Q_j'(s’, \pi'(s’;\theta’);\phi’_j) – \min_j Q_j(s,a;\phi_j)\). 9. Store transition \((s,a,r,s’,\delta)\) in SumTree with priority \(p = |\delta| + \epsilon\). 10. Sample a batch using SumTree priorities. 11. Compute importance sampling weights. 12. Update Critic networks by minimizing weighted MSE loss. 13. Every \(d\) steps, update Actor using sampled policy gradient. 14. Soft update target networks: \(\theta’ \leftarrow \tau \theta + (1-\tau)\theta’\), \(\phi’ \leftarrow \tau \phi + (1-\tau)\phi’\). 15. Update priorities in SumTree based on new TD errors. |
In my network design, I augment the standard multilayer perceptron with a Long Short-Term Memory layer and a self-attention layer. The Actor network takes the state as input, passes it through an LSTM layer with 128 units, then two fully connected layers with 256 units and ReLU activation, and finally outputs an action using Tanh activation. The Critic network concatenates the state and action, processes the state through a self-attention layer, then uses similar layers to produce a Q-value. This architecture improves temporal modeling and attention to critical environmental cues.
4.6 Experimental Setup and Results
I performed simulations in two three-dimensional environments: a large-obstacle environment (10 km × 10 km × 6 km) with spheres, cones, and cylinders, and a small-obstacle environment (15 km × 15 km × 3 km) with more obstacles. The main training parameters are listed in Table 4.
| Parameter | Value |
|---|---|
| Discount factor \(\gamma\) | 0.99 |
| Actor learning rate | 0.001 |
| Critic learning rate | 0.001 |
| Soft update coefficient \(\tau\) | 0.005 |
| Priority coefficient \(\alpha_p\) | 0.6 |
| Importance sampling exponent \(\beta\) | 0.4 |
| Experience replay capacity | 10^6 |
| Batch size | 512 |
| Max steps per episode | 500 |
| Max episodes | 500 |
I compared SPER-TD3 with eight baseline algorithms: MADDPG, fully decentralized TD3, DDPG, SAC, PPO, and centralized variants. The evaluation metrics include trajectory length, total reward, success rate, safe trajectory coverage, and an aggregated importance score. The importance score is computed by normalizing each metric and applying equal weights:
$$I = \frac{1}{4}(L’ + R’ + S’ + C’)$$
where \(L’\), \(R’\), \(S’\), and \(C’\) are the normalized values of trajectory length, reward, success rate, and safe coverage, respectively.
Table 5 presents the comparative results.
| Algorithm | Trajectory length | Reward magnitude | Success rate (%) | Safe coverage (%) | Importance score |
|---|---|---|---|---|---|
| SPER-TD3 | 13.26 | 29.96 | 98.80 | 33.33 | 0.650 |
| MADDPG | 13.17 | 29.92 | 99.00 | 19.40 | 0.596 |
| Decentralized TD3 | 13.33 | 30.92 | 98.40 | 24.84 | 0.585 |
| Decentralized DDPG | 12.84 | 28.54 | 97.40 | 22.72 | 0.599 |
| Decentralized SAC | 12.74 | 29.86 | 92.20 | 30.43 | 0.478 |
| Centralized SAC | 13.77 | 31.47 | 90.60 | 26.57 | 0.498 |
| Centralized TD3 | 13.60 | 30.70 | 98.00 | 31.43 | 0.642 |
| Centralized DDPG | 13.78 | 32.01 | 94.00 | 31.22 | 0.477 |
| Centralized PPO | 13.92 | 31.91 | 72.20 | 26.57 | 0.376 |
My proposed method achieves the highest importance score of 0.650, indicating superior overall performance. In addition, SPER-TD3 converges faster and exhibits more stable training curves than the baselines in both environments. The trajectory smoothing algorithm effectively reduces the steering angles of followers, as shown in Table 6, which compares average turn angles before and after smoothing.
| UAV | Follower 1 (before) | Follower 1 (after) | Follower 2 (before) | Follower 2 (after) |
|---|---|---|---|---|
| Average turn angle (deg) | 89.61 | 51.67 | 114.03 | 54.68 |
The results confirm that SPER-TD3 enhances trajectory safety, smoothness, and formation robustness in complex three-dimensional environments.
5. Interesting Region Selection and Transmission Fusion for Bandwidth-Limited UAVs
5.1 Motivation
Unmanned aerial vehicles often face severe bandwidth constraints when transmitting video to the ground control station. Transmitting full-resolution video at high frame rates consumes excessive bandwidth and may cause packet loss and frame drops. To overcome this, I propose a region-of-interest (ROI) selection and fusion method that identifies high-value regions in real time and transmits only those regions, along with low-frequency static background frames, to the ground station.
5.2 Method Overview
The proposed method consists of several stages. First, the ground station specifies target object categories and the flight path. The unmanned aerial vehicle then captures video and measures the available transmission bandwidth. Onboard object detection using YOLOv10 is performed on each frame. If a target from the specified categories is detected, the unmanned aerial vehicle enters a hovering state and focuses its gimbal on the target. After the video stabilizes, an ROI is generated around the target. The ROI video is then prioritized for transmission based on the measured bandwidth and the importance of the target categories. At the ground station, the received ROI video is fused with the last stored static background frame to reconstruct a visually complete scene. Figure 5.1 (not shown) illustrates the workflow.
Algorithm 5.1 outlines the dynamic ROI processing logic.
| Algorithm: Dynamic ROI Processing |
|---|
|
1. Initialize: target indices, static ROI marker = None, movement threshold = 50 pixels, max static time = 2 seconds. 2. For each frame in video stream: 3. Detect objects and keep boxes with confidence > 0.5 and target classes. 4. If a valid detection exists: 5. Compute the center of the largest box. 6. If static ROI exists and current center is inside it: 7. Increment static time. 8. If static time ≥ 2 seconds: capture the static ROI region. 9. Else: 10. Update dynamic ROI to the current box region; reset static time. 11. Else: 12. Clear static ROI marker. 13. If a captured static ROI exists: 14. Create background from the previously saved static frame. 15. Overlay the current ROI visual result onto the background. 16. Else: 17. Use the current dynamic ROI. 18. Output the fused ROI. |
5.3 TensorRT Acceleration
Deploying YOLOv10 on an edge computing platform requires aggressive optimization. I utilize NVIDIA TensorRT to accelerate inference. TensorRT applies layer fusion, precision calibration (FP16/INT8), dynamic tensor memory management, and kernel auto-tuning. The inference pipeline is as follows: the trained ONNX model is parsed and optimized into an engine file; input frames are preprocessed (resized, normalized, and converted to RGB); the engine runs inference; and post-processing outputs bounding boxes and class probabilities.
I conducted experiments on a Jetson AGX Orin platform using two scenarios: a sparse crowd scene and a dense crowd scene. Table 7 compares the average inference time and frame rate under CPU, GPU, and TensorRT modes.
| Computing environment | Inference time (ms) – Sparse | Inference time (ms) – Dense | FPS – Sparse | FPS – Dense |
|---|---|---|---|---|
| CPU | 224.23 | 224.37 | 4.48 | 4.48 |
| GPU | 88.82 | 87.89 | 14.48 | 14.51 |
| TensorRT | 28.48 | 27.86 | 45.71 | 46.15 |
The TensorRT-optimized model achieves an average inference speed of 28.17 ms and a frame rate of 45.93 FPS, which is a 217% improvement over the GPU baseline and fully meets real-time requirements.
5.4 Bandwidth Reduction Results
In the dense crowd experiment, the ROI-based transmission reduced the peak video bandwidth by 50% and the average bandwidth by 35.2%, while keeping the missed detection rate below 2%. Figure 5.2 (not shown) demonstrates that the transmitted ROI video retains clear details of the target region, while the background is refreshed at a low rate. This approach significantly alleviates wireless link pressure and improves the reliability of real-time video transmission for unmanned aerial vehicles.
6. Design and Validation of an Edge Computing Object Detection Software
6.1 Software Architecture
I designed a complete object detection software system for unmanned aerial vehicle edge computing scenarios. The system consists of a ground station component and an onboard component. The ground station uses a PyQt5-based graphical user interface for video display, remote command sending via SSH, and bandwidth analysis. The onboard component runs on the Jetson AGX Orin and includes the TensorRT-accelerated YOLOv10 detector, the ROI selection module, an RTSP server, and a wireless bandwidth measurement module.
The software flowchart is as follows:
- Unmanned aerial vehicles complete trajectory planning based on the SPER-TD3 algorithm.
- Onboard cameras capture video and feed it to the edge computing platform.
- The TensorRT model performs object detection.
- The ROI algorithm extracts critical regions.
- An RTSP server streams the ROI video to the ground station.
- The ground station displays the ROI video and bandwidth usage curves.
6.2 Interface Design
The target detection interface contains four modules: ROI settings, ROI display, video stream settings, and bandwidth comparison. The user can select target classes, set ROI parameters, configure the RTSP address, and start or stop the video stream. The ROI display panel shows the received ROI video, and the bandwidth comparison panel plots real-time bandwidth curves for both the original and ROI streams.
The bandwidth testing interface includes server/client parameter settings, real-time bandwidth and latency graphs, and a result recording module. It allows the user to configure IP addresses, ports, protocols, test duration, and bitrate. The interface visualizes throughput and latency over time, enabling comprehensive performance evaluation of the wireless link.
6.3 Validation in a Real Scenario
I validated the software using a real unmanned aerial vehicle platform (DJI M350 RTK) with a H30T gimbal camera and Ubinode self-organizing network modules. The edge device was a Jetson AGX Orin. The wireless bandwidth test showed an average throughput of 505.2 Mbits/sec with negligible jitter, confirming a high-quality link. In the field experiment, the software successfully detected and displayed the ROI around a crowd of spectators while the rest of the frame remained as a static background. The ROI bandwidth was significantly lower than the full video stream, demonstrating the effectiveness of the proposed system in real-world conditions. The experiment confirmed that the software is robust and practical for deployment in traffic management and military reconnaissance scenarios.
7. Conclusion and Future Work
In this research, I have addressed three major challenges for unmanned aerial vehicle formations: (1) stable and smooth trajectory planning in dense three-dimensional environments, (2) bandwidth-efficient video transmission through ROI selection, and (3) practical edge computing software for real-time detection.
My contributions are as follows. First, I proposed the SPER-TD3 algorithm that integrates SumTree prioritized experience replay with an improved TD3 architecture. The method employs an angle-based virtual leader formation controller and a dynamic trajectory smoothing scheme. Experiments show that SPER-TD3 outperforms eight mainstream deep reinforcement learning algorithms in terms of trajectory length, reward, success rate, safe coverage, and convergence stability. Second, I developed a dynamic ROI selection and fusion method that reduces video transmission bandwidth by up to 50% while preserving critical information. Third, I designed and implemented an edge computing object detection software using TensorRT and PyQt5, which achieves high-speed inference (46 FPS) and robust performance in real-world tests.
Future works may extend the trajectory planning method to heterogeneous formations, incorporate multimodal sensor data for ROI extraction under low-visibility conditions, and integrate more advanced lightweight detectors into the edge software. The ultimate goal is to enable fully autonomous unmanned aerial vehicle swarms that can adapt to dynamic environments and make intelligent decisions in real time.