A multi-camera target recognition and tracking system for unmanned aerial vehicles

Abstract

The application discloses a kind of unmanned aerial vehicle multi-camera target identification and tracking system, belong to unmanned aerial vehicle technical field, the unmanned aerial vehicle multi-camera target identification and tracking system includes a rear-end tracking unit and multiple front-end image processing units connected therewith;The front-end image processing unit includes image acquisition and dynamic pre-processing module, improved YOLOv8n target detection and feature extraction module and PTP synchronous data packaging transmission module;The rear-end tracking unit includes PTP dynamic time bucket synchronization module, parameter adaptive cascade tracking module and global resource dynamic scheduling module;The present disclosure solves the problem of existing system time synchronization difference, weak shielding robustness, low spatial accuracy, applicable to unmanned aerial vehicle high-altitude reconnaissance scene.

Description

Technical Field

[0001] This application belongs to the field of unmanned aerial vehicle (UAV) technology, and in particular relates to a multi-camera target recognition and tracking system for UAVs. Background Technology

[0002] With the development of drone reconnaissance technology, multi-camera collaborative target tracking has become a core requirement. Existing technologies suffer from the following key shortcomings: 1. Insufficient time synchronization accuracy: Data from multiple cameras is often aligned using simple timestamps without considering differences in transmission delay, resulting in a global frame synthesis deviation of >5ms, causing tracking gaps; 2. Poor robustness in occlusion scenarios: Single tracking algorithms (such as ByteTrack which relies solely on motion matching and DeepSORT which fuses appearance features with fixed weights) exhibit an ID switching rate >30% when targets are cross-occluded. 3. Rigid edge resource scheduling: Edge chips (such as RV1126B) are deployed using a fixed model, and there is no tiered degradation strategy when the NPU is overloaded, resulting in a sharp drop in detection frame rate of >40%; 4. Weak small target detection capability: For small targets of 20×20 pixels in high-altitude scenes, the existing YOLO series models have a detection rate of less than 60%, which is difficult to meet the reconnaissance requirements.

[0003] Existing technologies mostly focus on optimizing a single module and have not formed a complete solution of "edge preprocessing - time synchronization - cascaded tracking - spatial fusion - dynamic scheduling". Furthermore, they have not designed dedicated mechanisms for high-altitude small targets and dynamic computing scenarios of UAVs, leaving room for improvement.

[0004] Therefore, it is necessary to provide a new multi-camera target recognition and tracking system for drones to solve the above-mentioned technical problems. Summary of the Invention

[0005] The purpose of this disclosure is to provide a multi-camera target recognition and tracking system for unmanned aerial vehicles (UAVs) in order to solve the above-mentioned problems.

[0006] This disclosure achieves the above objectives through the following technical solutions: A multi-camera target recognition and tracking system for unmanned aerial vehicles (UAVs) includes a back-end tracking unit and multiple front-end image processing units connected thereto. The front-end image processing unit includes an image acquisition and dynamic preprocessing module, an improved YOLOv8n target detection and feature extraction module, and a PTP synchronous data packaging and transmission module. The image acquisition and dynamic preprocessing module is used to acquire raw images, perform preprocessing, and implement dynamic RGA scaling. The improved YOLOv8n target detection and feature extraction module is used to perform target detection and feature extraction using an improved model. The PTP synchronous data packaging and transmission module is used to package and transmit the processed data. The backend tracking unit includes a PTP dynamic time bucket synchronization module, a parameter adaptive cascaded tracking module, and a global resource dynamic scheduling module. The PTP dynamic time bucket synchronization module is used to establish dynamic time buckets and adjust the bucket span according to the target's PTP timestamp deviation. The parameter adaptive cascaded tracking module is used to adaptively perform ByteTrack motion matching and weighted fusion StrongSORT appearance matching to maintain tracking consistency in occluded scenarios. The global resource dynamic scheduling module is used for heterogeneous resource binding and hierarchical degradation strategy generation.

[0007] As a further optimization of this disclosure, the image acquisition and dynamic preprocessing module preprocesses the original image by: dynamically adjusting the noise reduction intensity, HDR exposure time, and color correction matrix according to the light intensity.

[0008] As a further optimization of this disclosure, the specific method for the image acquisition and dynamic preprocessing module to implement dynamic RGA scaling is as follows: if the proportion of small targets in the previous frame is greater than a preset percentage, the scaling ratio is set to 4K to 720P; otherwise, it is set to 4K to 1080P.

[0009] As a further optimization of this disclosure, the improved YOLOv8n object detection and feature extraction module is constructed by adding a small object enhancement branch to the neck of the YOLOv8n model. The small object enhancement branch adopts a combination of 3×3 convolution and 1×1 convolution and embeds an attention mechanism.

[0010] As a further optimization of this disclosure, the data packet structure of the PTP synchronization data packet transmission module adopts a header-data segment-checksum format. The header contains the camera ID, frame ID, and PTP timestamp; the data segment contains the number of targets, the bounding box of each target, the category, the confidence level, and the ReID feature; and the checksum uses CRC32.

[0011] As a further optimization of this disclosure, the PTP synchronization data packet transmission module adopts batch transmission and priority scheduling, and packages a data packet for every preset number of frames. If the USB buffer occupancy is greater than the preset percentage, target data with a confidence level less than the preset value is discarded first.

[0012] As a further optimization of this disclosure, the time synchronization process implemented by the PTP dynamic time bucket synchronization module includes: Receive data packets, extract the PTP timestamp of the data packets, calculate the timestamp deviation, dynamically adjust the time bucket span based on the timestamp deviation, and mark data in the same bucket as global frames; if a certain data path is missing more than a preset number of frames, predict the target position of the current frame based on the trajectory of the previous preset number of frames.

[0013] As a further optimization of this disclosure, the adaptive ByteTrack motion matching and weighted fusion StrongSORT appearance matching of the parameter adaptive cascaded tracking module include: Adjust the noise covariance Q of the Kalman filter process based on the target velocity v; adjust the IoU matching threshold based on the target density; Motion similarity S_m and appearance similarity S_a are calculated, and the fusion weight ω is adjusted according to the degree of occlusion. The final similarity S = ω × S_m + (1-ω) × S_a. An improved ResNet-34 ReID model is used to output 128-dimensional features.

[0014] As a further optimization of this disclosure, the heterogeneous resource allocation method of the global resource dynamic scheduling module is as follows: Bind the tracing thread to CPU cores 0-1 and the data receiving thread to cores 2-3.

[0015] As a further optimization of this disclosure, it also includes a full-link anomaly handling mechanism: when the camera disconnects, reconnection and trajectory prediction are initiated; when the NPU is overloaded, the YOLOv8n-tiny model is switched or ReID is turned off; and when the USB is blocked, low-confidence target data is discarded first.

[0016] The beneficial effects of this disclosure are as follows: The front-end achieves accurate detection of small targets through dynamic RGA scaling and improved YOLOv8n, while embedding PTP timestamps to ensure data synchronization. The back-end employs dynamic time bucket synchronization to improve the temporal accuracy of multiple cameras, reduces the ID switching rate in occlusion scenarios through cascaded tracking of "adaptive ByteTrack + weighted StrongSORT," and designs a global resource hierarchical degradation strategy. This invention solves the problems of poor time synchronization, weak occlusion robustness, and low spatial accuracy in existing systems, and is suitable for UAV high-altitude reconnaissance scenarios. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is an overall structural framework diagram of an embodiment of this disclosure; Figure 2 This is a flowchart of the front-end image processing unit in an embodiment of this disclosure; Figure 3 This is a logic diagram of the dynamic RGA scaling algorithm in an embodiment of this disclosure.

[0019] Figure 4 This is a flowchart of the parameter adaptive cascaded tracking algorithm in an embodiment of this disclosure; Figure 5 This is a flowchart of the system exception handling in an embodiment of this disclosure. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] like Figure 1 As shown, a multi-camera target recognition and tracking system for unmanned aerial vehicles includes eight RV1126B image processing units at the front end and one RK3588 tracking unit at the back end. The specific scheme is as follows: (a) Front-end RV1126B image processing unit, such as Figure 2 As shown: 1. Image acquisition and dynamic preprocessing module: MIPI capture: Raw images captured from the IMX415 4K camera at 30 FPS; ISP adaptive processing: dynamically adjusts noise reduction intensity (noise reduction coefficient 1.2-1.5 for low-light scenes, 0.8-1.0 for strong light scenes), HDR exposure time (50-200μs), and color correction matrix according to light intensity; Dynamic RGA scaling: such as Figure 3As shown, a "target size perception algorithm" is introduced. If the proportion of small targets (≤20×20 pixels) in the previous frame is >20%, the scaling ratio is set to 4K→720P (to preserve the details of small targets); otherwise, it is set to 4K→1080P (to reduce computing power consumption). The image size after scaling is fed back to the algorithm layer in real time through a circular buffer.

[0022] 2. Improved YOLOv8n target detection and feature extraction module: Model Improvement: A "small target enhancement branch" is added to the YOLOv8n neck. This branch uses a combination of 3×3 convolutions and 1×1 convolutions to increase the receptive field coverage (from the original 32×32 to 16×16) and embeds an attention mechanism (SE module) to enhance the response of small target features. Model training: The "UAV high-altitude dataset" (including categories such as people and vehicles, with small target samples accounting for 35%) was used. Data augmentation was achieved through Mosaic (stitching ratio 1:3), random scaling (0.8-1.2 times), and grayscale perturbation (±10%). The PyTorch framework was used with an initial learning rate of 1e-3, a cosine annealing strategy, a batch size of 16, and 120 training rounds. Quantization deployment: After exporting the ONNX model, INT8 quantization was performed using RKNN Toolkit2 (the calibration dataset is 10% of the dataset samples). After quantization, the model parameters are 1.7M, memory usage is ≤80MB, and NPU inference time is ≤30ms.

[0023] 3. PTP Synchronous Data Packet Transmission Module: Data packet structure design: The "header-data segment-checksum" format is adopted. The header contains camera ID (1 byte), frame ID (4 bytes), and PTP timestamp (8 bytes, precision 1ns); the data segment contains the number of targets (2 bytes), the bounding box of each target (4×4 bytes, x1 / y1 / x2 / y2), category (1 byte), confidence level (1 byte, quantized to 0-255), and 128-dimensional ReID features (128 bytes, INT8 quantization); the checksum uses CRC32 (4 bytes). USB 3.0 transmission optimization: "Batch transmission + priority scheduling" is adopted, and one data packet is packaged every 2 frames (reducing transmission overhead). If the USB buffer usage is >85%, target data with confidence <0.5 is discarded first to ensure the transmission of critical targets.

[0024] (ii) Back-end RK3588 tracking unit 1. PTP Dynamic Time Bucket Synchronization Module: Time synchronization process: After receiving 8 RV1126B data packets, extract the PTP timestamp and establish a "dynamic time bucket" (initial time span of 10ms, if the delay of a certain data is >5ms, the bucket span is automatically expanded to 15ms); mark the 8 data packets whose timestamps fall in the same bucket as "global frames", with a synchronization accuracy of ≤1ms; Abnormal data handling: If more than 3 frames of data are missing from a certain channel, the "completion mechanism" is triggered. The target position in the current frame is predicted based on the target motion trajectory of the previous 3 frames to avoid global frame loss.

[0025] 2. Parameter adaptive cascaded tracking module, such as Figure 4 As shown: This module employs a two-level cascaded structure: the first level is adaptive ByteTrack motion matching, and the second level is weighted fusion StrongSORT appearance matching. The two levels are linked by data association logic to ensure tracking consistency even in occluded scenarios. The specific implementation steps are as follows: Level 1: Adaptive ByteTrack Motion Matching Step 2.1: Target State Prediction Kalman filtering is used to predict the state of each tracked target. The state vector includes position (x, y), velocity (v_x, v_y), and acceleration (a_x, a_y).

[0026] The dynamic adjustment process noise covariance Q is calculated based on the target motion speed v (pixels / frame), where v = √(v_x² + v_y²). If v > 50, then Q = 0.15I (identity matrix) to accommodate fast motion; if v ≤ 50, then Q = 0.05I to improve smoothness.

[0027] Prediction equation: X_pred=F×X_prev, where F is the state transition matrix, designed based on a uniform velocity model.

[0028] Step 2.2: IoU Matching and Threshold Adaptation Calculate the IoU (Intersection over Union) between the detected bounding box and the predicted bounding box.

[0029] Dynamically adjust the IoU threshold: based on the target density D (number of targets / frame) of the current frame. If D>20, set the threshold to 0.25 to reduce false matches; if D≤20, set the threshold to 0.35 to improve the matching rate.

[0030] Matching process: First, IoU matching is performed on high-confidence detection boxes, then a second matching is performed on low-confidence detection boxes, and unmatched trackers are temporarily retained.

[0031] Step 2.3: Output the matching results Output successfully matched tracker-detection box pairs, as well as unmatched trackers and detection boxes, for secondary processing.

[0032] Level 2: Weighted Fusion StrongSORT Appearance Matching Step 2.4: ReID Feature Extraction ReID features for each bounding box were extracted using an improved ResNet-34 model (with the last two convolutional layers removed, outputting 128-dimensional features). The model was INT8 quantized, and NPU inference time was ≤5ms. Feature extraction was performed on unmatched bounding boxes and trackers.

[0033] Step 2.5: Similarity Calculation and Weighted Fusion Motion similarity S_m: Based on IoU distance transformation, S_m=1-IoU_distance.

[0034] Appearance similarity S_a: based on cosine distance transformation, S_a=cosine_similarity(ReID_features).

[0035] The fusion weight ω is dynamically adjusted based on the target occlusion rate O (calculated through pixel overlap rate). If O ≤ 30%, ω = 0.7 (emphasizing motion); if O > 30%, ω = 0.3 (emphasizing appearance).

[0036] The final similarity S = ω × S_m + (1-ω) × S_a.

[0037] Step 2.6: Data Association and LSTM Smoothing The Hungarian algorithm was used for matching based on similarity S, with a threshold of 0.5.

[0038] LSTM smoothing is applied to the target position (x,y) of three consecutive frames: the LSTM model input is a historical coordinate sequence, the number of hidden layer nodes is 32, and the output is the smoothed coordinates to reduce vibration noise.

[0039] 3. Global resource dynamic scheduling module: The CPU A76×4 cores are bound to critical threads, with the tracing thread bound to cores 0-1 (priority SCHED_FIFO=99) and the data receiving thread bound to cores 2-3 (priority SCHED_FIFO=90); the NPU is dedicated to ReID feature extraction; the Mali-G610 optimizes homography matrix calculation through OpenCL.

[0040] The first stage quickly associates targets through motion matching, while the second stage handles occlusion through appearance matching. Weighted fusion ensures strong adaptability. LSTM smoothing is used as post-processing to improve trajectory continuity. The entire process is a closed-loop process; unmatched targets are initialized as new trackers, and lost targets are deleted after lifecycle management.

[0041] (iii) End-to-end anomaly handling mechanism, such as Figure 5 As shown:

[0042] The hardware setup for this system is as follows: I. Hardware Architecture Setup Front-end hardware: 8 RV1126B core boards (model RV1126B-DDR4-2GB), each core board is connected to one IMX415 4K camera (resolution 3840×2160, frame rate 30FPS), and connected to the back-end via USB3.0 Type-C interface; the core board is powered by 12V / 2A, and the power consumption is ≤5W; Backend hardware: RK3588 core board (model RK3588-DDR4-8GB, NPU computing power 6TOPS INT8), equipped with USB 3.0 HUB (supports 8 simultaneous accesses), communicates with the ground station via Ethernet (1000Mbps); the core board is powered by 12V / 3A, and the power consumption is ≤10W; Drone integration: The eight cameras adopt a "mask-like layout" (circularly distributed around the nose, covering a 360° field of view, with adjacent cameras overlapping by 15°). The front core board is centrally installed in the middle of the fuselage, and the rear core board is installed at the tail of the fuselage to avoid electromagnetic interference.

[0043] II. Software Module Implementation Steps (I) Front-end RV1126B software implementation (based on Linux 4.19 kernel) 1. Image acquisition and preprocessing module MIPI driver configuration: Using the official Rockchip MIPI driver, the acquisition resolution is set to 3840×2160 and the frame rate is 30FPS; ISP parameter configuration: Set the noise reduction mode to "multi-frame noise reduction" and the HDR mode to "dynamic range 12EV" using the RKISP2 tool; the color correction matrix adopts the sRGB standard. Dynamic RGA scaling: Based on the libRGA library, a scaling algorithm is written. For each frame, the proportion of small targets is first detected (by statistical analysis of the detection results from the previous frame), and then the RGA interface is called to set the scaling ratio. The scaling time is ≤2ms.

[0044] 2. Improved YOLOv8n target detection and feature extraction module Model training: On an Ubuntu 20.04 system, in a PyTorch 2.0 environment, a custom "UAV high-altitude dataset" (100,000 frames, including 5,000 small target samples) was used to train the improved YOLOv8n model, and the ONNX model was exported after training was completed; Model quantization: Using RKNN Toolkit2, load the ONNX model, select the "Image Calibration" mode (calibration dataset is 200 frames of samples), generate the RKNN model, and the accuracy loss after quantization is ≤2%; Inference Deployment: Based on the librknnrt library, inference code was written to bind the NPU inference thread to RV1126B CPU core 1, with inference time ≤30ms.

[0045] 3. PTP Synchronous Data Packet Transmission Module PTP timestamp acquisition: Synchronize 8 RV1126B clocks with an accuracy of ≤1ns using the Linux PTP tool (ptp4l), and record the PTP timestamp when each frame of image is acquired; Data packet encapsulation: Implement the data packet format using C language structures, call the libusb library to implement USB 3.0 transmission, set up batch transmission endpoints, pack 1 data packet every 2 frames, and achieve a transmission rate of ≥400Mbps.

[0046] (II) Backend RK3588 software implementation (based on Linux 5.10 kernel) 1. PTP Dynamic Time Bucket Synchronization Module Dynamic time bucket creation: Write C++ code to receive 8 data packets, extract PTP timestamps, calculate timestamp deviations, dynamically adjust the time bucket span, and achieve synchronization time ≤3ms; The completion mechanism is implemented by predicting the target position of the missing frame based on Kalman filtering, with a prediction error of ≤10 pixels.

[0047] 2. Parameter Adaptive Cascaded Tracking Module Adaptive ByteTrack: Modify the ByteTrack open source code, add speed threshold judgment logic, dynamically adjust the Q matrix and IoU threshold, and the motion matching time is ≤4ms; Weighted StrongSORT: Based on the StrongSORT open-source framework, it integrates and improves the ReID model (RKNN format), writes weighted fusion code, and achieves appearance matching time ≤6ms; LSTM smoothing: Deploy an LSTM model (input dimension 2, hidden layer 32 nodes, output dimension 2) using TensorFlow Lite, and smoothing time ≤ 2ms.

[0048] 3. Global Resource Dynamic Scheduling Module Thread binding: Use the pthread library to bind the tracing thread to CPU cores 0-1 and the data receiving thread to cores 2-3; Load monitoring: CPU load is monitored using the sysmon tool, and NPU load is monitored using rknn_drv. Data is collected every 50ms, and each module is notified via semaphores when degradation is triggered.

[0049] III. System Testing and Performance Verification Test environment: The drone flew at an altitude of 100m and a speed of 15m / s. The scene contained 20 targets (including 5 small targets of 20×20 pixels) with an occlusion rate of 30%. Test metrics: Time synchronization accuracy: 0.8ms; End-to-end latency: 75ms; Tracking frame rate: 18 FPS; ID switching rate: 7.2%; Small target detection rate: 89%; 3D coordinate error: 7.5%; Frame rate after NPU overload degradation: 15 FPS (meets minimum requirements).

[0050] The embodiments described above are merely examples of several implementations of this disclosure, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent disclosure. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this disclosure, and these modifications and improvements all fall within the protection scope of this disclosure.

Claims

1. A multi-camera target recognition and tracking system for unmanned aerial vehicles (UAVs), characterized in that, It includes a back-end tracking unit and multiple front-end image processing units connected to it; The front-end image processing unit includes an image acquisition and dynamic preprocessing module, an improved YOLOv8n target detection and feature extraction module, and a PTP synchronous data packaging and transmission module. The image acquisition and dynamic preprocessing module is used to acquire raw images, perform preprocessing, and implement dynamic RGA scaling. The improved YOLOv8n target detection and feature extraction module is used to perform target detection and feature extraction using an improved model. The PTP synchronous data packaging and transmission module is used to package and transmit the processed data. The backend tracking unit includes a PTP dynamic time bucket synchronization module, a parameter adaptive cascaded tracking module, and a global resource dynamic scheduling module. The PTP dynamic time bucket synchronization module is used to establish dynamic time buckets and adjust the bucket span according to the target's PTP timestamp deviation. The parameter adaptive cascaded tracking module is used to adaptively perform ByteTrack motion matching and weighted fusion StrongSORT appearance matching to maintain tracking consistency in occluded scenarios. The global resource dynamic scheduling module is used for heterogeneous resource binding and hierarchical degradation strategy generation.

2. The UAV multi-camera target recognition and tracking system according to claim 1, characterized in that, The image acquisition and dynamic preprocessing module preprocesses the original image by dynamically adjusting the noise reduction intensity, HDR exposure time, and color correction matrix according to the light intensity.

3. The UAV multi-camera target recognition and tracking system according to claim 1, characterized in that, The specific method for the image acquisition and dynamic preprocessing module to implement dynamic RGA scaling is as follows: if the proportion of small targets in the previous frame is greater than the preset percentage, the scaling ratio is set to 4K to 720P; otherwise, it is set to 4K to 1080P.

4. The UAV multi-camera target recognition and tracking system according to claim 1, characterized in that, The improved YOLOv8n object detection and feature extraction module is constructed by adding a small object enhancement branch to the neck of the YOLOv8n model. The small object enhancement branch uses a combination of 3×3 convolution and 1×1 convolution and embeds an attention mechanism.

5. The UAV multi-camera target recognition and tracking system according to claim 1, characterized in that, The data packet structure of the PTP synchronization data packet transmission module adopts a header-data segment-checksum format. The header contains the camera ID, frame ID, and PTP timestamp; the data segment contains the number of targets, the bounding box of each target, the category, the confidence level, and the ReID feature. The check segment uses CRC32.

6. The UAV multi-camera target recognition and tracking system according to claim 1, characterized in that, The PTP synchronization data packet transmission module adopts batch transmission and priority scheduling. It packages a data packet for every preset number of frames. If the USB buffer usage is greater than the preset percentage, the target data with a confidence level less than the preset value is discarded first.

7. The UAV multi-camera target recognition and tracking system according to claim 1, characterized in that, The time synchronization process implemented by the PTP dynamic time bucket synchronization module includes: Receive data packets, extract the PTP timestamp of the data packets, calculate the timestamp deviation, dynamically adjust the time bucket span based on the timestamp deviation, and mark data in the same bucket as global frames; if a certain data path is missing more than a preset number of frames, predict the target position of the current frame based on the trajectory of the previous preset number of frames.

8. The UAV multi-camera target recognition and tracking system according to claim 1, characterized in that, The parameter adaptive cascaded tracking module includes adaptive ByteTrack motion matching and weighted fusion StrongSORT appearance matching, which include: Adjust the noise covariance Q of the Kalman filter process based on the target velocity v; adjust the IoU matching threshold based on the target density; Motion similarity S_m and appearance similarity S_a are calculated, and the fusion weight ω is adjusted according to the degree of occlusion. The final similarity S = ω × S_m + (1-ω) × S_a. An improved ResNet-34 ReID model is used to output 128-dimensional features.

9. A multi-camera target recognition and tracking system for unmanned aerial vehicles according to claim 1, characterized in that, The heterogeneous resource allocation method of the global resource dynamic scheduling module is as follows: Bind the tracing thread to CPU cores 0-1 and the data receiving thread to cores 2-3.

10. A multi-camera target recognition and tracking system for unmanned aerial vehicles according to claim 1, characterized in that, It also includes a full-link anomaly handling mechanism: when the camera disconnects, it initiates reconnection and trajectory prediction; when the NPU is overloaded, it triggers YOLOv8n-tiny model switching or ReID shutdown; and when the USB is blocked, it prioritizes discarding low-confidence target data.

Images