A low-altitude unmanned aerial vehicle autonomous tracking system based on computer vision

Abstract

The application relates to a low-altitude unmanned aerial vehicle autonomous tracking system based on computer vision, relates to the unmanned aerial vehicle tracking and sensing technical field, and comprises a multi-modal adaptive sensing module, an intelligent identification and relocation module, a three-dimensional state accurate estimation module, a predictive trajectory planning module, a multi-loop coupling intelligent control module and an edge energy efficiency optimization management module. The low-altitude unmanned aerial vehicle autonomous tracking system based on computer vision significantly improves the all-weather operation capability of the system in a complex environment by integrating the multi-modal adaptive sensing module and the intelligent identification and relocation module, utilizes multi-source data fusion of visible light, infrared, laser radar and millimeter wave radar, the system can effectively overcome the interference of rain, fog, night and strong light backlight and other bad weather on a single sensor, a time sequence memory unit based on a Transformer architecture is introduced, the system has the ability to remember the historical trajectory of a target, and can quickly find back the target in a millisecond level time by combining inertial calculation and a multi-hypothesis search strategy.

Description

Technical Field

[0001] This invention relates to the field of drone tracking and perception technology, specifically to a low-altitude drone autonomous tracking system based on computer vision. Background Technology

[0002] With the booming development of the low-altitude economy and the widespread application of UAV technology, computer vision-based autonomous tracking systems have become one of the core capabilities of UAVs in performing reconnaissance, search and rescue, and logistics tasks. Existing low-altitude UAV autonomous tracking systems typically adopt the classic perception-decision-control architecture: at the perception layer, the system mainly relies on monocular or binocular visible light cameras to acquire target images, with some high-end devices supplemented by simple infrared sensors; at the algorithm layer, traditional feature point matching (such as SIFT, ORB) or early convolutional neural networks (CNN) are mostly used for target detection and tracking, combined with Kalman filtering to estimate the target's position; at the control layer, the UAV's attitude is adjusted through a PID controller to maintain its relative position. These systems perform reasonably well in scenarios with good lighting, simple backgrounds, and linear target motion, and can achieve basic following functions, and are widely used in consumer-grade aerial photography and simple industrial inspections.

[0003] However, existing computer vision-based autonomous tracking systems for low-altitude UAVs have significant shortcomings in complex real-world application scenarios. First, single-modal perception methods lack robustness; in all-weather conditions such as rain, fog, nighttime, or strong light / backlight, visible light cameras are prone to failure, leading to target loss and difficulty in rapid retrieval. Second, traditional algorithms lack effective temporal memory mechanisms; once the target is obscured by buildings, trees, etc., the system often loses all connection due to its inability to predict the target's trajectory, resulting in time-consuming relocation processes and a high risk of false detections. Third, existing systems mostly rely on static or simplified motion models for state estimation, making it difficult to accurately calculate the target's 3D depth pose and six-degree-of-freedom dynamic changes, leading to safety hazards when constructing collision avoidance boundaries. Furthermore, existing trajectory planning and control strategies are mostly open-loop or simple closed-loop, lacking the ability to predict the target's future movement trends, resulting in uneven paths that can easily cause severe UAV shaking or even loss of control. Finally, existing systems suffer from rigid computing power scheduling, failing to dynamically balance computing resources according to task stages, making it difficult to simultaneously meet the requirements of low-latency real-time response and long endurance on edge devices. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a low-altitude UAV autonomous tracking system based on computer vision, which has advantages such as multi-source heterogeneous data fusion and solves the problem of target loss in adverse environments such as rain, fog, and night.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a low-altitude unmanned aerial vehicle autonomous tracking system based on computer vision, comprising a multimodal adaptive perception module, an intelligent identification and relocation module, a three-dimensional state accurate estimation module, a predictive trajectory planning module, a multi-loop coupled intelligent control module, and an edge energy efficiency optimization management module; The multimodal adaptive perception module, used to fuse multi-source data to achieve all-weather robust capture, includes a visible light camera, an infrared thermal imager, a lidar, and a millimeter-wave radar. The visible light camera and the infrared thermal imager are aligned at the frame level through a spatiotemporal synchronization unit. The lidar is used to acquire high-precision point cloud depth information, and the millimeter-wave radar is used to penetrate rain and fog environments to acquire the radial velocity of the target. The multimodal adaptive perception module also includes an image enhancement neural network, which is configured to receive ambient light intensity sensor data in real time. When the light intensity is lower than a preset threshold or the contrast is lower than a set value, it automatically adjusts the exposure time, gain coefficient, and defogging algorithm parameters to map the multi-source data to a unified world coordinate system to generate a 3D voxel feature map. The intelligent recognition and relocation module is used to memorize temporal trajectories and quickly retrieve targets in the event of occlusion. It includes a target detection network based on the Transformer architecture, a temporal memory unit, and a multi-hypothesis search strategy engine. The temporal memory unit adopts a long short-term memory network or a gated recurrent unit and is configured to store the trajectory sequence features of the target over the past N frames, where N ranges from 5 to 50. When the target is occluded and the confidence level is lower than a first preset threshold, the multi-hypothesis search strategy engine uses the coordinates of the occluded area calculated by the inertial measurement unit and combines them with historical motion vectors to perform a rapid scan within the local field of view. After detecting matching features, it triggers an online fine-tuning mechanism to update the target recognition model to adapt to specific appearance features. The three-dimensional state accurate estimation module is used to calculate depth pose and construct dynamic collision avoidance boundaries. It includes a monocular depth estimation network, an extended Kalman filter, and a dynamic safety boundary constructor. The monocular depth estimation network is configured to output the relative depth map of the target and perform weighted fusion with the lidar point cloud data to calculate the target's absolute position coordinates and six-degree-of-freedom pose on the X, Y, and Z axes. The extended Kalman filter is configured to update the target's velocity and acceleration vectors at a frequency of 50Hz to 200Hz. The dynamic safety boundary constructor dynamically calculates and maintains a virtual ellipsoidal safety envelope that changes with the target's maneuverability based on the current relative distance and the predicted collision time. When the UAV enters this envelope, an emergency braking command is triggered. The predictive trajectory planning module is used to predict motion trends and generate smooth obstacle avoidance paths. It includes a nonlinear motion predictor and a model prediction controller. The nonlinear motion predictor is configured to predict the target's position trajectory within the next T seconds based on the target's state vector at the current moment, where T ranges from 0.5 seconds to 2.0 seconds. The model prediction controller is configured to, based on the predicted trajectory, combine the UAV's kinematic constraint equations and obstacle avoidance map, with the objective function of minimizing tracking error and energy consumption, to generate a smooth reference waypoint sequence through rolling optimization, and output control commands including longitudinal velocity, lateral velocity, and yaw rate. The multi-loop coupled intelligent control module, used for dual closed-loop coordinated action to eliminate jitter and achieve precise tracking, includes a gimbal attitude inner loop controller, a flight control displacement outer loop controller, and a kinematic compensation unit. The gimbal attitude inner loop controller is configured to drive the gimbal motor at millisecond-level frequencies to eliminate high-frequency vibrations of the fuselage and keep the target centered in the field of view. The flight control displacement outer loop controller is configured to adjust the UAV rotor speed according to the reference trajectory output by the predictive trajectory planning module to maintain a preset relative distance and angle. The kinematic compensation unit is configured to preset a reverse compensation amount before the control command is issued, based on the UAV's own accelerometer and gyroscope data, to offset the tracking lag error caused by inertial torque. The edge energy efficiency optimization management module is used to dynamically schedule computing power and balance low latency and long battery life. It includes a dynamic computing power scheduler, a lightweight model inference engine, and breakpoint resume logic. The dynamic computing power scheduler is configured to dynamically allocate computing resources according to the task stage. In search mode, it reduces the image processing resolution and frame rate to 15fps, and in locked mode, it increases the resolution to over 30fps and runs the high-precision model at full speed. The lightweight model inference engine deploys a neural network model that has undergone knowledge distillation and INT8 quantization to reduce memory usage and improve inference speed. The breakpoint resume logic is configured to automatically switch to a pure local offline mode when the network signal is interrupted, continue to execute the tracking task by relying on the pre-trained model, and upload key log data after the signal is restored.

[0006] Furthermore, the system also includes a communication link management unit, which is used to automatically switch to a 5G / 4G cellular network or a private image transmission link when the visible light communication link is blocked, and to compress and encode the pose data calculated by the three-dimensional state accurate estimation module and the path data generated by the predictive trajectory planning module and transmit them to the ground station to realize remote monitoring and manual takeover.

[0007] Furthermore, the system also includes an environmental perception assistance unit, which is integrated around the UAV fuselage and includes an ultrasonic ranging sensor and a stereo vision sensor. This unit is used to supplement the data blind spots of the multimodal adaptive perception module in complex low-altitude environments. When a static obstacle in front is detected to be less than a preset safety threshold, the detour logic of the predictive trajectory planning module is triggered first.

[0008] Furthermore, the multi-source data fusion in the multimodal adaptive perception module adopts a feature fusion algorithm based on an attention mechanism. The attention mechanism is configured to dynamically adjust the fusion ratio of visible light, infrared and radar data according to the signal-to-noise ratio weight of each sensor, so as to ensure that the target tracking success rate of the system is not less than 95% under strong light, backlight or bad weather conditions.

[0009] Furthermore, the communication link management unit is also configured to encrypt the compressed and encoded data to ensure the security of data transmission.

[0010] Furthermore, the sampling frequencies of the ultrasonic ranging sensor and the stereo vision sensor are synchronized with the processing cycle of the multimodal adaptive sensing module.

[0011] Furthermore, the object detection network based on the Transformer architecture employs a multi-head attention mechanism to extract global features and combines spatial convolution operations to improve the recognition accuracy of local features.

[0012] Furthermore, the state transition matrix of the extended Kalman filter includes dynamic model parameters, and the covariance matrix is ​​adaptively adjusted according to the current observation noise level.

[0013] Furthermore, the kinematic constraint equations used by the model predictive controller take into account the maximum rotational angular velocity and maximum linear velocity limits of the UAV, as well as the dynamic response delay of the onboard battery.

[0014] Furthermore, the input node dimension of the neural network model deployed by the lightweight model inference engine is adapted to different data stream formats, and hidden layers that may cause memory overflow or invalid computation are dynamically pruned according to the current available memory size.

[0015] Compared with the prior art, the technical solution of this application has the following beneficial effects: 1. This computer vision-based low-altitude UAV autonomous tracking system significantly improves the system's all-weather operation capability in complex environments by integrating a multimodal adaptive perception module and an intelligent recognition and relocation module. Utilizing multi-source data fusion from visible light, infrared, lidar, and millimeter-wave radar, the system can effectively overcome the interference of adverse weather conditions such as rain, fog, nighttime, and strong light / backlight on a single sensor, achieving robust target acquisition. At the same time, the introduction of a temporal memory unit based on the Transformer architecture enables the system to remember the target's historical trajectory. Once the target is occluded, it can quickly retrieve the target within milliseconds by combining inertial calculation and multi-hypothesis search strategies, completely solving the technical bottlenecks of easy loss and difficult relocation in traditional monocular vision tracking.

[0016] 2. This computer vision-based low-altitude UAV autonomous tracking system achieves high-precision and high-stability low-altitude autonomous tracking through the synergistic effect of four modules: accurate 3D state estimation, predictive trajectory planning, multi-loop coupled intelligent control, and edge energy efficiency optimization management. The dynamic collision avoidance boundary constructed by the 3D state estimation module effectively ensures flight safety; the predictive trajectory planning module generates a smooth obstacle avoidance path by predicting the target's movement trend, eliminating the lag effect in traditional control; the multi-loop coupled intelligent control module significantly suppresses high-frequency vibrations of the fuselage through coordinated inner and outer loop actions, ensuring a smooth and accurate following process; and the edge energy efficiency optimization management module dynamically schedules computing power according to the task, significantly reducing power consumption while ensuring low latency and extending the UAV's endurance. The combined effect of these technical features gives the system a higher tracking success rate, stronger anti-interference capability, and better energy utilization when dealing with high-speed maneuvering targets and complex dynamic scenarios. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 The low-altitude UAV autonomous tracking system based on computer vision in this embodiment includes a multimodal adaptive perception module, an intelligent identification and relocation module, a three-dimensional state accurate estimation module, a predictive trajectory planning module, a multi-loop coupled intelligent control module, and an edge energy efficiency optimization management module. The multimodal adaptive perception module, used to fuse multi-source data to achieve robust all-weather capture, includes a visible light camera, an infrared thermal imager, a lidar, and a millimeter-wave radar. The visible light camera and the infrared thermal imager are aligned at the frame level through a spatiotemporal synchronization unit. The lidar is used to acquire high-precision point cloud depth information, and the millimeter-wave radar is used to penetrate rain and fog environments to acquire the radial velocity of the target. The multimodal adaptive perception module also includes an image enhancement neural network, which is configured to receive ambient light sensor data in real time. When the light intensity is lower than a preset threshold or the contrast is lower than a set value, it automatically adjusts the exposure time, gain coefficient, and dehazing algorithm parameters to map the multi-source data to a unified world coordinate system to generate a 3D voxel feature map. The intelligent recognition and relocalization module is used to memorize temporal trajectories and quickly retrieve targets in the event of occlusion. It includes a target detection network based on the Transformer architecture, a temporal memory unit, and a multi-hypothesis search strategy engine. The temporal memory unit adopts a long short-term memory network or a gated recurrent unit and is configured to store the trajectory sequence features of the target over the past N frames, where N ranges from 5 to 50. When the target is occluded and the confidence level is lower than a first preset threshold, the multi-hypothesis search strategy engine uses the coordinates of the occluded area calculated by the inertial measurement unit and combines them with historical motion vectors to perform a fast scan within the local field of view. After detecting matching features, it triggers an online fine-tuning mechanism to update the target recognition model to adapt to specific appearance features. The 3D state accurate estimation module is used to calculate depth pose and construct dynamic collision avoidance boundaries. It includes a monocular depth estimation network, an extended Kalman filter, and a dynamic safety boundary builder. The monocular depth estimation network is configured to output the relative depth map of the target and perform weighted fusion with the lidar point cloud data to calculate the target's absolute position coordinates and six-degree-of-freedom pose on the X, Y, and Z axes. The extended Kalman filter is configured to update the target's velocity and acceleration vectors at a frequency of 50Hz to 200Hz. The dynamic safety boundary builder dynamically calculates and maintains a virtual ellipsoidal safety envelope that changes with the target's maneuverability based on the current relative distance and the predicted collision time. When the UAV enters this envelope, an emergency braking command is triggered. The predictive trajectory planning module is used to predict motion trends and generate smooth obstacle avoidance paths. It includes a nonlinear motion predictor and a model predictive controller. The nonlinear motion predictor is configured to predict the target's position trajectory within the next T seconds based on the target's state vector at the current moment, where T ranges from 0.5 seconds to 2.0 seconds. The model predictive controller is configured to generate a smooth reference waypoint sequence through rolling optimization based on the predicted trajectory, combined with the UAV's kinematic constraint equations and obstacle avoidance map, with the objective function of minimizing tracking error and energy consumption, and output control commands including longitudinal velocity, lateral velocity, and yaw rate. The multi-loop coupled intelligent control module is used for dual closed-loop coordinated action to eliminate jitter and achieve precise tracking. It includes a gimbal attitude inner loop controller, a flight control displacement outer loop controller, and a kinematic compensation unit. The gimbal attitude inner loop controller is configured to drive the gimbal motors at millisecond-level frequencies to eliminate high-frequency vibrations of the fuselage and keep the target centered in the field of view. The flight control displacement outer loop controller is configured to adjust the UAV rotor speed according to the reference trajectory output by the predictive trajectory planning module to maintain a preset relative distance and angle. The kinematic compensation unit is configured to preset a reverse compensation amount before the control command is issued based on the UAV's own accelerometer and gyroscope data to offset the tracking lag error caused by inertial torque. The edge energy efficiency optimization management module is used to dynamically schedule computing power and balance low latency and long battery life. It includes a dynamic computing power scheduler, a lightweight model inference engine, and breakpoint resume logic. The dynamic computing power scheduler is configured to dynamically allocate computing resources according to the task stage. In search mode, it reduces the image processing resolution and frame rate to 15fps, and in locked mode, it increases the resolution to over 30fps and runs the high-precision model at full speed. The lightweight model inference engine deploys a neural network model that has undergone knowledge distillation and INT8 quantization to reduce memory usage and improve inference speed. The breakpoint resume logic is configured to automatically switch to a pure local offline mode when the network signal is interrupted, continue to execute the tracking task by relying on the pre-trained model, and upload key log data after the signal is restored.

[0020] An autonomous tracking system for low-altitude unmanned aerial vehicles (UAVs) comprises six modules (multimodal perception, intelligent recognition, 3D estimation, trajectory planning, multi-loop control, and energy efficiency management). Each module integrates specific hardware combinations, algorithm architectures (such as Transformer, EKF, and MPC), and control logic to achieve all-weather robust capture, occlusion relocalization, accurate pose calculation, predictive obstacle avoidance, and efficient operation at the edge. The system also includes a communication link management unit, which automatically switches to a 5G / 4G cellular network or a private image transmission link when the visible light communication link is blocked. It also compresses and encodes the pose data calculated by the three-dimensional state accurate estimation module and the path data generated by the predictive trajectory planning module and transmits them to the ground station to achieve remote monitoring and manual takeover.

[0021] It should be noted that a communication link management unit has been added, which supports automatic switching to cellular or image transmission network when visible light is blocked, and compresses the transmitted pose and path data to support remote monitoring.

[0022] The system also includes an environmental perception auxiliary unit, which is integrated around the drone's fuselage. This unit includes ultrasonic ranging sensors and stereo vision sensors, used to supplement the data blind spots of the multimodal adaptive perception module in complex low-altitude environments. When a static obstacle ahead is detected to be less than a preset safety threshold, the detour logic of the predictive trajectory planning module is triggered first.

[0023] It should be noted that an environmental perception auxiliary unit (ultrasound and stereo vision) has been added to supplement low-altitude blind spots and to prioritize triggering detour logic when a nearby static obstacle is detected.

[0024] Among them, the multi-source data fusion in the multimodal adaptive perception module adopts a feature fusion algorithm based on the attention mechanism. The attention mechanism is configured to dynamically adjust the fusion ratio of visible light, infrared and radar data according to the signal-to-noise ratio weight of each sensor, so as to ensure that the target tracking success rate of the system is not less than 95% under strong light, backlight or bad weather conditions.

[0025] It should be noted that the multimodal perception module is limited to an attention-based feature fusion algorithm, which dynamically adjusts the sensor weights according to the signal-to-noise ratio to ensure a high success rate in adverse weather conditions.

[0026] The communication link management unit is also configured to encrypt the compressed and encoded data to ensure the security of data transmission.

[0027] It should be noted that the communication link management unit is required to encrypt the transmitted data to ensure data transmission security.

[0028] The sampling frequencies of the ultrasonic ranging sensor and the stereo vision sensor are synchronized with the processing cycle of the multimodal adaptive sensing module.

[0029] It should be noted that the sampling frequency of the environmental perception auxiliary unit must be synchronized with the processing cycle of the main perception module.

[0030] Among them, the object detection network based on the Transformer architecture uses a multi-head attention mechanism to extract global features and combines spatial convolution operations to improve the recognition accuracy of local features.

[0031] It should be noted that the specific target detection network adopts a multi-head attention mechanism combined with spatial convolution operation to balance global feature extraction and local recognition accuracy.

[0032] The extended Kalman filter's state transition matrix contains dynamic model parameters, and its covariance matrix is ​​adaptively adjusted based on the current observation noise level.

[0033] It should be noted that the state transition matrix of the specifically defined extended Kalman filter includes dynamic parameters, and the covariance matrix can be adaptively adjusted according to the observation noise.

[0034] The kinematic constraint equations used by the model predictive controller take into account the maximum rotational angular velocity and maximum linear velocity limits of the UAV, as well as the dynamic response delay of the onboard battery.

[0035] It should be noted that the constraint equations for the specific model predictive controller take into account the maximum speed limit of the UAV and the battery power response delay.

[0036] The lightweight model inference engine adapts the input node dimensions of the neural network model to different data stream formats and dynamically trims hidden layers that may cause memory overflow or invalid computation based on the available memory size.

[0037] It should be noted that the neural network input dimension of the specific lightweight inference engine is configurable, and the hidden layers can be dynamically pruned according to the memory size to optimize resource consumption.

[0038] The working principle of the above embodiment is as follows: by constructing a temporal memory unit based on the Transformer architecture, the target features of multiple historical frames are input in time sequence, and the long-distance dependency relationship is directly captured by the self-attention mechanism to establish a global motion model. When the target is occluded and visual features are lost, the system automatically generates a smooth future prediction path as a priori guidance based on the learned historical trajectory topology, thereby achieving millisecond-level fast relocalization and seamless tracking without the need for re-searching.

[0039] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A low-altitude unmanned aerial vehicle (UAV) autonomous tracking system based on computer vision, characterized in that, It includes a multimodal adaptive perception module, an intelligent recognition and relocation module, a three-dimensional state accurate estimation module, a predictive trajectory planning module, a multi-loop coupled intelligent control module, and an edge energy efficiency optimization management module; The multimodal adaptive perception module, used to fuse multi-source data to achieve all-weather robust capture, includes a visible light camera, an infrared thermal imager, a lidar, and a millimeter-wave radar. The visible light camera and the infrared thermal imager are aligned at the frame level through a spatiotemporal synchronization unit. The lidar is used to acquire high-precision point cloud depth information, and the millimeter-wave radar is used to penetrate rain and fog environments to acquire the radial velocity of the target. The multimodal adaptive perception module also includes an image enhancement neural network, which is configured to receive ambient light intensity sensor data in real time. When the light intensity is lower than a preset threshold or the contrast is lower than a set value, it automatically adjusts the exposure time, gain coefficient, and defogging algorithm parameters to map the multi-source data to a unified world coordinate system to generate a 3D voxel feature map. The intelligent recognition and relocation module is used to memorize temporal trajectories and quickly retrieve them in the event of occlusion. It includes a target detection network based on the Transformer architecture, a temporal memory unit, and a multi-hypothesis search strategy engine. The temporal memory unit adopts a long short-term memory network or a gated recurrent unit and is configured to store the trajectory sequence features of the target over the past N frames, where the value of N ranges from 5 to 50. When the target is occluded and the confidence level is lower than the first preset threshold, the multi-hypothesis search strategy engine uses the coordinates of the occluded area calculated by the inertial measurement unit, combined with the historical motion vector, to perform a fast scan within the local field of view, and triggers an online fine-tuning mechanism after detecting matching features to update the target recognition model to adapt to specific appearance features. The three-dimensional state accurate estimation module is used to calculate depth pose and construct dynamic collision avoidance boundaries. It includes a monocular depth estimation network, an extended Kalman filter, and a dynamic safety boundary constructor. The monocular depth estimation network is configured to output the relative depth map of the target and perform weighted fusion with the lidar point cloud data to calculate the target's absolute position coordinates and six-degree-of-freedom pose on the X, Y, and Z axes. The extended Kalman filter is configured to update the target's velocity and acceleration vectors at a frequency of 50Hz to 200Hz. The dynamic safety boundary constructor dynamically calculates and maintains a virtual ellipsoidal safety envelope that changes with the target's maneuverability based on the current relative distance and the predicted collision time. When the UAV enters this envelope, an emergency braking command is triggered. The predictive trajectory planning module is used to predict motion trends and generate smooth obstacle avoidance paths. It includes a nonlinear motion predictor and a model prediction controller. The nonlinear motion predictor is configured to predict the target's position trajectory within the next T seconds based on the target's state vector at the current moment, where T ranges from 0.5 seconds to 2.0 seconds. The model prediction controller is configured to, based on the predicted trajectory, combine the UAV's kinematic constraint equations and obstacle avoidance map, with the objective function of minimizing tracking error and energy consumption, to generate a smooth reference waypoint sequence through rolling optimization, and output control commands including longitudinal velocity, lateral velocity, and yaw rate. The multi-loop coupled intelligent control module, used for dual closed-loop coordinated action to eliminate jitter and achieve precise tracking, includes a gimbal attitude inner loop controller, a flight control displacement outer loop controller, and a kinematic compensation unit. The gimbal attitude inner loop controller is configured to drive the gimbal motor at millisecond-level frequencies to eliminate high-frequency vibrations of the fuselage and keep the target centered in the field of view. The flight control displacement outer loop controller is configured to adjust the UAV rotor speed according to the reference trajectory output by the predictive trajectory planning module to maintain a preset relative distance and angle. The kinematic compensation unit is configured to preset a reverse compensation amount before the control command is issued, based on the UAV's own accelerometer and gyroscope data, to offset the tracking lag error caused by inertial torque. The edge energy efficiency optimization management module is used to dynamically schedule computing power and balance low latency and long battery life. It includes a dynamic computing power scheduler, a lightweight model inference engine and breakpoint resume logic. The dynamic computing scheduler is configured to dynamically allocate computing resources according to the task stage. In search mode, it reduces the image processing resolution and frame rate to 15fps, and in locked mode, it increases the resolution to over 30fps and runs the high-precision model at full speed. The lightweight model inference engine deploys a neural network model that has undergone knowledge distillation and INT8 quantization to reduce memory usage and improve inference speed. The breakpoint resume logic is configured to automatically switch to pure local offline mode when the network signal is interrupted, continue to execute the tracking task by relying on the pre-trained model, and upload key log data after the signal is restored.

2. The low-altitude unmanned aerial vehicle (UAV) autonomous tracking system based on computer vision according to claim 1, characterized in that: The system also includes a communication link management unit, which automatically switches to a 5G / 4G cellular network or a private image transmission link when the visible light communication link is blocked. It also compresses and encodes the pose data calculated by the three-dimensional state accurate estimation module and the path data generated by the predictive trajectory planning module and transmits them to the ground station to realize remote monitoring and manual takeover.

3. The low-altitude unmanned aerial vehicle (UAV) autonomous tracking system based on computer vision according to claim 1, characterized in that: The system also includes an environmental perception assistance unit, which is integrated around the UAV fuselage and includes an ultrasonic ranging sensor and a stereo vision sensor. It is used to supplement the data blind spots of the multimodal adaptive perception module in complex low-altitude environments. When a static obstacle in front is detected to be less than a preset safety threshold, the detour logic of the predictive trajectory planning module is triggered first.

4. The low-altitude unmanned aerial vehicle (UAV) autonomous tracking system based on computer vision according to claim 1, characterized in that: The multi-source data fusion in the multimodal adaptive perception module adopts a feature fusion algorithm based on an attention mechanism. The attention mechanism is configured to dynamically adjust the fusion ratio of visible light, infrared and radar data according to the signal-to-noise ratio weight of each sensor, so as to ensure that the target tracking success rate of the system is not less than 95% under strong light, backlight or bad weather conditions.

5. The low-altitude unmanned aerial vehicle (UAV) autonomous tracking system based on computer vision according to claim 1, characterized in that: The communication link management unit is also configured to encrypt the compressed and encoded data to ensure the security of data transmission.

6. The low-altitude unmanned aerial vehicle autonomous tracking system based on computer vision according to claim 1, characterized in that: The sampling frequencies of the ultrasonic ranging sensor and the stereo vision sensor are synchronized with the processing cycle of the multimodal adaptive sensing module.

7. The low-altitude unmanned aerial vehicle autonomous tracking system based on computer vision according to claim 1, characterized in that: The object detection network based on the Transformer architecture uses a multi-head attention mechanism to extract global features and combines spatial convolution operations to improve the recognition accuracy of local features.

8. The low-altitude unmanned aerial vehicle autonomous tracking system based on computer vision according to claim 1, characterized in that: The state transition matrix of the extended Kalman filter contains dynamic model parameters, and the covariance matrix is ​​adaptively adjusted according to the current observation noise level.

9. The low-altitude unmanned aerial vehicle (UAV) autonomous tracking system based on computer vision according to claim 1, characterized in that: The kinematic constraint equations used by the model predictive controller take into account the maximum rotational angular velocity and maximum linear velocity limits of the UAV, as well as the dynamic response delay of the onboard battery.

10. A low-altitude unmanned aerial vehicle (UAV) autonomous tracking system based on computer vision according to claim 1, characterized in that: The lightweight model inference engine deploys neural network models whose input node dimensions are adapted to different data stream formats, and dynamically trims hidden layers that may cause memory overflow or invalid computation based on the available memory size.

Images