Real-time scene collision detection method and system for low-altitude unmanned aerial vehicle

Abstract

The invention discloses a real-time scene collision detection method and system for a low-altitude unmanned aerial vehicle, and relates to the technical field of unmanned aerial vehicles, and the method comprises the steps: selecting an optimal obstacle avoidance action according to a classification result of objects in a scene, generating a flight path through employing a path planning algorithm, and meanwhile, in the flight process, carrying out the optimal obstacle avoidance action; the state of the unmanned aerial vehicle and environment change are monitored in real time for collision detection, and meanwhile, an obstacle avoidance strategy is adjusted. Through the decision-making method based on the comprehensive reward value, the safety of the unmanned aerial vehicle is improved, smooth completion of the flight task is ensured, the energy consumption is optimized, the endurance time of the unmanned aerial vehicle is prolonged, in addition, the unmanned aerial vehicle continuously learns and optimizes the obstacle avoidance strategy according to the historical flight data, and the obstacle avoidance efficiency is improved. And the autonomous flight capability and efficiency are further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles, and in particular to a real-time scene collision detection method and system for low-altitude unmanned aerial vehicles. Background Art

[0002] Low-altitude drones have developed rapidly in recent years and are widely used in many fields such as aerial photography, logistics distribution, and agricultural monitoring. As the application scope of drones expands, their flight environment is becoming increasingly complex, especially in low-altitude scenes, where drones face challenges from many dynamic and static obstacles. In order to ensure the safety of drones, real-time scene collision detection technology has become a key element in autonomous flight of drones.

[0003] Existing technologies still have shortcomings in low-altitude complex scenes. First, low-altitude drones fly at high speeds, and changes in the environment require obstacle avoidance actions to respond within milliseconds. However, existing obstacle avoidance methods usually rely on complex computing models, resulting in large processing delays and failure to meet real-time requirements. Especially when facing multiple obstacles, the drone's response speed will be further reduced, increasing the risk of collision. Summary of the invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a real-time scene collision detection method for low-altitude unmanned aerial vehicles to solve the problem of insufficient real-time performance of existing obstacle avoidance methods.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a real-time scene collision detection method for a low-altitude UAV, which includes using a variety of sensors to collect data about the surrounding environment of the low-altitude UAV in real time when it is flying;

[0008] Fuse the collected data to generate a unified environmental model;

[0009] Classify objects in the scene based on the fused environment model;

[0010] Select the best obstacle avoidance action based on the classification results;

[0011] A path planning algorithm is used to generate a flight path based on the best obstacle avoidance action;

[0012] During the flight, the drone's status and environmental changes are monitored in real time to detect collisions and adjust obstacle avoidance strategies.

[0013] As a preferred solution of the real-time scene collision detection method for low-altitude UAVs of the present invention, a plurality of sensors are used to collect data of the surrounding environment of the low-altitude UAV in real time during flight, including the following steps:

[0014] Collect historical flight data of drones in real time;

[0015] The multiple sensors are laser radar, millimeter wave radar, camera, ultrasonic wave and GPS / IMU sensor;

[0016] Use a laser radar sensor to emit a laser beam to the surrounding area, extract the three-dimensional space coordinates according to the angle of the laser beam, and output point cloud data;

[0017] Use a millimeter-wave radar sensor to modulate the frequency of radar transmission and output radar data of the distance, speed and reflection intensity of the target point;

[0018] Using a camera sensor to capture an image through a lens and output image data;

[0019] Use ultrasonic sensors to transmit ultrasonic signals and receive echo signals, and output the distance data between the obstacle and the target point;

[0020] Use GPS / IMU sensors to output position information data and attitude angle data respectively.

[0021] As a preferred solution of the real-time scene collision detection method for low-altitude UAVs described in the present invention, the collected data is fused to generate a unified environment model, including the following steps:

[0022] Use statistical filtering algorithm to eliminate noise in point cloud data and perform sampling;

[0023] Use the constant false alarm rate detection method to extract the distance, speed and angle information of the target point in the radar data and generate a target point set;

[0024] De-noise and contrast enhance the image data and convert it into digital signals;

[0025] The distance data between obstacles and target points that exceed the effective detection distance are eliminated through the threshold value;

[0026] After processing, the boundaries, shape, and size of obstacles are extracted from the LiDAR sensor;

[0027] Extract the speed and direction of motion of dynamic obstacles from the millimeter wave radar sensor;

[0028] Extracting semantic and depth information from camera sensors;

[0029] Extracting distance information of close obstacles from ultrasonic sensors;

[0030] The extracted environmental features of LiDAR, millimeter-wave radar, camera, and ultrasound are integrated to form the state vector of the obstacle.

[0031] Based on the state vectors of obstacles, a unified environment model is generated.

[0032] As a preferred solution of the real-time scene collision detection method for low-altitude UAVs described in the present invention, objects in the scene are classified based on the fused environment model, including the following steps:

[0033] Select the convolutional neural network learning model and ReLU activation function;

[0034] The state vector of the obstacle is input into the activation function, and the features of the obstacle are extracted after linear transformation;

[0035] The obstacle features are linearly transformed again through the convolutional neural network learning model to obtain the classification probability distribution;

[0036] According to the classification probability distribution, all pixels of the object image are processed one by one, and the classification probability of each pixel is calculated. The expression is:

[0037]

[0038] Among them, C represents the classification probability of each pixel, Z represents the normalization constant, σ represents the normalization function that converts the category score into a probability distribution, W represents the weight matrix, φ(S) represents the new feature vector after being processed by the activation function, and b represents the bias term;

[0039] Based on the classification probability of each pixel, the maximum probability category of the object pixel is calculated and the category threshold of the object image pixel is set;

[0040] When the maximum probability of an object pixel is greater than or equal to the category threshold of the object image pixel, the object is marked as dynamic;

[0041] When the maximum probability of an object pixel is less than the category threshold of the object image pixel, the object is marked as static.

[0042] As a preferred solution of the real-time scene collision detection method for low-altitude UAVs described in the present invention, the best obstacle avoidance action is selected according to the classification result, including the following steps:

[0043] Use deep convolutional neural networks to perform semantic segmentation on images captured by the camera and generate semantic maps;

[0044] Based on the semantic map, the DeepSORT target tracking algorithm is used in combination with image data to extract the position and speed of dynamic obstacles and output the state vector of the dynamic obstacles.

[0045] Combine the semantic map and target tracking results to generate an environment feature vector;

[0046] According to the safety, mission completion and energy consumption of the drone during flight, the reward function is set and the comprehensive reward value is calculated. The expression is:

[0047] R=w 1 ×D+w 2 ×E+w 3 ×F;

[0048] Among them, R represents the comprehensive reward value obtained by the drone when performing the corresponding action in flight, D represents the safety reward of the drone in flight, E represents the task completion reward of the drone in flight, F represents the energy consumption reward of the drone in flight, and w 1 Represents the weight coefficient of security, w 2 Represents the weight coefficient of task completion, w 3 Represents the weight coefficient of energy consumption;

[0049] Based on the obtained comprehensive reward value, the Bellman equation is used to update the expected cumulative reward of the drone in the future and combined with the maximum value of the expected cumulative reward to determine the obstacle avoidance action with the best long-term performance.

[0050] As a preferred solution of the real-time scene collision detection method for low-altitude UAVs of the present invention, wherein: according to the optimal obstacle avoidance action, a path planning algorithm is used to generate a flight path, including the following steps:

[0051] Based on the optimal obstacle avoidance action, combined with the density of static obstacles, the behavior pattern of dynamic obstacles and the specific flight mission requirements, a safe action set covering feasible actions and a safe distance for the drone to fly are generated;

[0052] Define the UAV's dynamic constraints based on its physical limitations and motion characteristics;

[0053] Construct a three-dimensional environment grid based on the current position of the UAV, the target point, the environmental feature vector, and the dynamic constraints of the UAV, put the UAV into the three-dimensional environment network, and output the current position coordinates of the UAV;

[0054] Initialize the starting flight node and target node in the three-dimensional environment grid, and define the open list and the closed list;

[0055] The open list is a set of nodes to be searched, and the closed list is a set of nodes that have been searched;

[0056] The heuristic cost from the start node to the target node is calculated using the Euclidean distance, expressed as:

[0057]

[0058] Among them, h(n) represents the heuristic cost from the starting node to the target point, n represents the starting node, g represents the target node, and x n Indicates the coordinate of the starting node on the x-axis, y n Indicates the coordinate of the starting node on the y-axis, z n Indicates the coordinate of the starting node on the z-axis, x g Indicates the coordinate of the target node on the x-axis, y g Indicates the coordinate of the target node on the y-axis, z g Indicates the coordinate of the target node on the z-axis;

[0059] The dynamic constraint cost is obtained by accumulating the turning radius and speed limit in the dynamic constraint of the UAV;

[0060] Taking the environmental feature vector as a benchmark, the actual cost from the starting point to the start node in the three-dimensional environmental grid is obtained;

[0061] According to the heuristic cost from the start node to the target node, the dynamic constraint cost and the actual cost from the starting point to the start node, the total cost function of all nodes is obtained, which is expressed as:

[0062] f(g)=r(i)+h(n)+c(j);

[0063] Among them, f(g) represents the total cost function of all nodes, r(i) represents the actual cost from the starting point to the starting node in the three-dimensional environment grid, and c(j) represents the dynamic constraint cost;

[0064] Add the current position of the drone to the open list, and select the starting node with the smallest total cost function from the open list;

[0065] Select the A* algorithm to search the flight path. When the smallest starting node is equal to the target point, the search is successful and the flight path is output;

[0066] When the smallest starting node is not equal to the target point, the total cost function of the neighboring nodes is updated and added to the open list, the smallest starting node is moved to the closed list, and the operation is repeated until the flight path is found.

[0067] As a preferred solution of the real-time scene collision detection method for low-altitude UAVs of the present invention, the following steps are included: during the flight, the state of the UAV and environmental changes are monitored in real time to perform collision detection and adjust the obstacle avoidance strategy:

[0068] Use the Kalman filter to predict the position of obstacles in the future. The expression is:

[0069]

[0070] Among them, O represents the predicted position of the obstacle in the future, G represents the obstacle position at the current time t, V represents the obstacle speed at the current time t, Δt represents the predicted time step from the current time t to the future, and P represents the acceleration of the obstacle;

[0071] The predicted obstacle position information in the future is placed in the three-dimensional environment grid, and the straight-line distance between the current position of the drone and the obstacle position in the future is calculated. The expression is:

[0072]

[0073] Where T represents the straight-line distance between the current position of the drone and the position of the obstacle in the future, u represents the drone, m represents the obstacle, and x represents the distance between the current position of the drone and the obstacle in the future. u Indicates the coordinate of the drone on the x-axis, y u Indicates the coordinate of the drone on the y-axis, z u Indicates the coordinate of the drone on the z-axis, x m Represents the predicted position coordinates of the obstacle on the x-axis, y m Represents the predicted position coordinate of the obstacle on the y-axis, z m Indicates the predicted position coordinates of the obstacle on the z-axis;

[0074] When the straight-line distance between the current position of the drone and the position of the obstacle in the future is greater than the safe distance for the drone to fly, no collision occurs and the drone operates normally;

[0075] When the straight-line distance between the current position of the drone and the position of the obstacle in the future is less than or equal to the safe distance for the drone to fly, there is a possibility of collision. The A* algorithm is used again to search and plan a new flight path, which is sent to the flight controller to monitor environmental changes in real time. The flight direction, speed and altitude of the drone are adjusted according to actual needs. Repeat the operation to further adjust the flight path.

[0076] According to the results of collision detection, adjust the obstacle avoidance strategy.

[0077] In a second aspect, the present invention provides a real-time scene collision detection system for low-altitude UAVs, including a data acquisition module that uses a variety of sensors to collect data about the surrounding environment of the low-altitude UAV in real time when it is flying;

[0078] The model generation module integrates the collected data to generate a unified environmental model;

[0079] The object classification module classifies objects in the scene based on the fused environment model;

[0080] The obstacle avoidance selection module selects the best obstacle avoidance action based on the classification results;

[0081] The path generation module generates the flight path using the path planning algorithm according to the optimal obstacle avoidance action;

[0082] Adjust the strategy module to monitor the status of the drone and environmental changes in real time during flight to perform collision detection and adjust the obstacle avoidance strategy.

[0083] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the real-time scene collision detection method for low-altitude unmanned aerial vehicles as described in the first aspect of the present invention is implemented.

[0084] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the real-time scene collision detection method for a low-altitude unmanned aerial vehicle as described in the first aspect of the present invention.

[0085] The beneficial effects of the present invention are: combining semantic maps and target tracking results to generate environmental feature vectors, setting reward functions based on the safety, mission completion and energy consumption of the drone, updating the expected cumulative rewards obtained by the drone in the future through the Bellman equation, and selecting the best obstacle avoidance action. This decision-making method based on comprehensive reward values ​​not only improves the safety of the drone, but also ensures the smooth completion of the flight mission, and extends the drone's flight time by optimizing energy consumption. In addition, the drone continuously learns and optimizes obstacle avoidance strategies based on historical flight data, further improving the ability and efficiency of autonomous flight. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0087] Figure 1 This is a flow chart of the real-time scene collision detection method for low-altitude UAV in Example 1.

[0088] Figure 2 Schematic diagram of a real-time scene collision detection system for low-altitude UAV in Example 1. DETAILED DESCRIPTION

[0089] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0090] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0091] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0092] Example 1, reference Figure 1 and Figure 2 , which is the first embodiment of the present invention, provides a real-time scene collision detection method for a low-altitude UAV, comprising the following steps:

[0093] S1. Use a variety of sensors to collect real-time data about the surrounding environment of low-altitude drones.

[0094] The following steps are included:

[0095] Use Webots flight simulator (UAV simulation platform) to simulate various flight scenarios of UAV in a virtual environment, and record events that occur during the flight (such as collisions, path adjustments, etc.) to analyze the behavior patterns of UAVs, form historical flight data and save the flight log on a dedicated UAV data recorder;

[0096] The various sensors include LiDAR, millimeter wave radar, camera, ultrasonic, and GPS / IMU sensors;

[0097] Use a laser radar sensor to emit a laser beam to the surrounding area, extract the three-dimensional space coordinates according to the angle of the laser beam, and output point cloud data;

[0098] LiDAR sensors are used for high-precision detection of long-distance obstacles. The data sampling frequency is generally set to 10Hz-20Hz, and the specific frequency depends on the flight speed of the drone and the real-time requirements;

[0099] Use a millimeter-wave radar sensor to modulate the frequency of radar transmission and output radar data of the distance, speed and reflection intensity of the target point;

[0100] Millimeter-wave radar is used for all-weather dynamic obstacle detection, especially in low-visibility environments (such as rain, snow, and foggy weather);

[0101] Using a camera sensor to capture an image through a lens and output image data;

[0102] Camera sensors are usually used to identify complex environmental features (such as buildings, pedestrians, vehicles, etc.), and the data sampling frequency is generally set to 30FPS-60FPS;

[0103] Use ultrasonic sensors to transmit ultrasonic signals and receive echo signals, and output the distance data between the obstacle and the target point;

[0104] Select ultrasonic sensors with high accuracy and fast response (such as HC-SR04) for close-range obstacle detection. The data sampling frequency is generally set to 10Hz-50Hz.

[0105] Use GPS / IMU sensors to output position information data and attitude angle data respectively;

[0106] The GPS module outputs position information (longitude, latitude, altitude, and timestamp) every second, and the IMU module outputs attitude angles (pitch, roll, and yaw). The data sampling frequency is generally set to 100Hz-200Hz.

[0107] S2. Fuse the collected data to generate a unified environmental model.

[0108] The following steps are included:

[0109] Use statistical filtering algorithm to eliminate noise in point cloud data and perform sampling;

[0110] The statistical filtering algorithm is a filtering method based on probability theory and statistics. It is widely used in signal processing, drone flight, robot navigation, autonomous driving and other fields. Its core idea is to process the observed data through mathematical models, remove noise and outliers, and extract useful information. It is suitable for scenarios with noise or incomplete information.

[0111] Calculate the average distance between each point and its neighboring points. If the average distance between a point and its neighboring points is much larger than that between other points, the point is considered an outlier and is removed. When point cloud data comes from different timestamps, it is necessary to ensure that all point cloud data are aligned in the same coordinate system. All point cloud data can be aligned to the same reference system through rigid transformation (translation, rotation), and the sampling rate can be adjusted according to the point cloud density in different areas. For example, in areas with low point cloud density (such as flat surfaces), the number of points can be appropriately reduced to reduce the amount of calculation.

[0112] Use the constant false alarm rate detection method to extract the distance, speed and angle information of the target point in the radar data and generate a target point set;

[0113] A continuous wave signal whose frequency changes linearly with time is emitted, and the frequency offset between the received echo signal and the continuous wave signal is calculated based on the Doppler effect to obtain the distance to the target point;

[0114] When the target's speed exceeds the radar's measurement range, a high-resolution radar system combined with the Doppler effect is used for auxiliary measurement to obtain speed information;

[0115] Equipped with antenna technology, the relationship between the signal phase difference and the target angle in the radar data is calculated to obtain the angle information;

[0116] Constant false alarm rate detection is an algorithm for target detection, especially suitable for data processing of radar, sonar and other sensors. Its core principle is to adaptively adjust the detection threshold to cope with changes in background noise while keeping the false alarm rate (i.e. false alarm rate) constant.

[0117] De-noise and contrast enhance the image data and convert it into digital signals;

[0118] Use the binocular camera to calculate the pixel color difference of the image, generate a depth map and output it in the form of frames in .jpg or .png format, and the depth map is output in .tiff or .raw format;

[0119] The distance data between obstacles and target points that exceed the effective detection distance are eliminated through the threshold value;

[0120] Threshold value refers to a critical value used to distinguish different states or categories in signal processing, image processing, sensor data processing and other fields;

[0121] In sensor data processing, thresholds are used to detect abnormal events or trigger alarms. For example, a temperature sensor can be set with a threshold value, and when the temperature exceeds this value, an alarm will be triggered or cooling measures will be taken. Similarly, an ultrasonic sensor can be set with a threshold value to detect whether an object has undergone violent movement or collision;

[0122] After processing, the boundaries, shape, and size of obstacles are extracted from the LiDAR sensor;

[0123] Use the K-means algorithm to cluster points belonging to the same object. Each cluster after clustering represents an independent obstacle. For each clustered obstacle, use the convex hull algorithm to extract its outer contour.

[0124] PCA (Principal Component Analysis) is used to estimate the main axes and directions of obstacles, which helps to identify the shape characteristics of objects (such as cuboids, cylinders, etc.);

[0125] Calculate the oriented bounding box of the obstacle. The size of the bounding box can directly reflect the size of the obstacle. Extract the speed and direction of movement of the dynamic obstacle from the millimeter wave radar sensor.

[0126] The Doppler map (two-dimensional matrix) is generated according to the Doppler effect. The speed of the target can be determined by finding the peak point from the Doppler map. By using dual-slope frequency modulation (up-chirp and down-chirp), the forward and reverse Doppler frequencies are measured respectively, and the radial speed and lateral speed of the target are separated, thereby estimating its direction of movement.

[0127] Extracting semantic and depth information from camera sensors;

[0128] Use target detection algorithms (YOLO, Faster R-CNN) to detect specific targets in the image and determine their location, size, and category; use depth sensors to emit infrared light and measure the time difference of reflected light to calculate the distance of the object and extract it;

[0129] Semantic information refers to the object category and its location information extracted from the image data. It not only includes the shape and color of the object, but more importantly, it can identify the specific category of the object (such as pedestrians, vehicles, buildings, trees, etc.);

[0130] For drone obstacle avoidance, semantic information can help drones identify which objects need to be avoided (such as pedestrians and vehicles) and which objects can be ignored (such as the sky and grass). For example, in city streets, semantic information can mark roads as "passable areas", pedestrians as "dynamic obstacles", and buildings as "static obstacles".

[0131] Depth information refers to the distance information of each point in an object or scene relative to the observer (such as a camera or lidar sensor);

[0132] Synchronize the timestamp of the obstacle state vector and perform feature matching of the obstacle;

[0133] Static obstacles: Associating the LiDAR point cloud data with the camera’s depth information, such as matching the building boundary point cloud detected by the LiDAR with the building outline in the camera image to determine the shape and specific semantic information of the obstacle;

[0134] Dynamic obstacles: associate the targets detected by the millimeter-wave radar sensor with the targets in the camera sensor through their target positions and motion trajectories, while keeping the distance, speed, and direction (millimeter-wave radar) of the target consistent with the depth and pixel position (camera);

[0135] The extracted data from LiDAR, millimeter-wave radar, camera, and ultrasound are integrated to form the state vector of the obstacle.

[0136] Based on the state vector of the obstacle, a three-dimensional environment map of each obstacle is constructed and updated using geometric modeling to form a unified environment model;

[0137] Unity refers to temporal unity, spatial unity, and semantic unity.

[0138] S3. Classify objects in the scene based on the fused environment model.

[0139] The following steps are included:

[0140] Select the convolutional neural network learning model and ReLU activation function;

[0141] The obstacle features are linearly transformed again through the convolutional neural network learning model to obtain the classification probability distribution;

[0142] The obstacle is characterized by the image and point cloud data acquired by the sensor;

[0143] According to the classification probability distribution, all pixels of the object image are processed one by one, and the classification probability of each pixel is calculated. The expression is:

[0144]

[0145] Among them, C represents the classification probability of each pixel. For example, a pixel may belong to a category such as pedestrians, vehicles, buildings, etc. The value of C represents the possibility that the pixel belongs to a certain category. For each pixel, it is finally assigned to a certain category according to the maximum probability. Z represents the normalization constant, which is used to make an overall adjustment to the calculation results so that all values ​​of the classification probability C of each pixel satisfy the properties of the probability distribution. σ represents the normalization function that converts the category score into the probability distribution. W represents the weight matrix, which is used to weight the different dimensions of the feature. φ(S) represents the new feature vector after the activation function is processed. Its function is to introduce nonlinear characteristics to remove the influence of negative values ​​and retain positive values, thereby enhancing the expression ability of the model. S represents the state vector of the obstacle, including multidimensional features such as depth information, speed information, and semantic information. b represents the bias term, which is used to translate the result of the linear transformation to enhance the flexibility of the model.

[0146] Based on the classification probability of each pixel, the maximum probability category of the object pixel is calculated as the final classification result of the object pixel. The expression is:

[0147] I = argmax(C);

[0148] Among them, I represents the maximum probability category of the object pixel;

[0149] Set the category threshold of the object image pixels. When the maximum probability of the object pixels is greater than or equal to the category threshold of the object image pixels, the object is marked as dynamic.

[0150] When the maximum probability of an object pixel is less than the category threshold of the object image pixel, the object is marked as static;

[0151] For example, assuming that the category threshold of object image pixels is 0.6, there are four object image pixels of "road", "pedestrian", "vehicle" and "building", and the corresponding classification probabilities of each pixel are 0.15, 0.7, 0.63 and 0.4. The maximum probability category of object pixels is "pedestrian". The probability of "pedestrian" object image pixels 0.7 is greater than the category threshold of object image pixels 0.6, so "pedestrian" is classified into the dynamic category.

[0152] S4. Select the best obstacle avoidance action based on the classification results.

[0153] The following steps are included:

[0154] Use deep convolutional neural networks to perform semantic segmentation on images captured by the camera and generate semantic maps;

[0155] Based on the semantic map, the DeepSORT target tracking algorithm is used in combination with image data to extract the position and speed of dynamic obstacles and output the state vector of the dynamic obstacles.

[0156] The DeepSORT target tracking algorithm is an algorithm widely used in multi-target tracking tasks, especially in video surveillance, autonomous driving, drone obstacle avoidance and other fields. It combines deep learning and traditional Kalman filtering technology to track multiple targets stably and accurately in a real-time environment.

[0157] Combine the semantic map and target tracking results to generate an environment feature vector;

[0158] According to the safety, mission completion and energy consumption of the drone during flight, the reward function is set and the comprehensive reward value is calculated. The expression is:

[0159] R=w 1 ×D+w 2 ×E+x 3 ×F;

[0160] Among them, R represents the comprehensive reward value obtained by the drone when performing the corresponding action in the flight state, which is used to measure the performance of the drone at the current moment. The larger the comprehensive reward value, the better the current decision (action) of the drone. D represents the safety reward of the drone in the flight state. If the drone successfully avoids obstacles, D gives a positive reward. If the drone collides with obstacles, D gives a negative reward. E represents the task completion reward of the drone in the flight state. If the drone approaches the target point according to the planned path, E gives a positive reward. If the drone deviates from the path or moves away from the target point, E gives a negative reward. F represents the energy consumption reward of the drone in the flight state. If the energy consumption of the drone action is low, F gives a positive reward. If the energy consumption of the drone action is high, F gives a negative reward. w 1 Represents the weight coefficient of security, w 2 Represents the weight coefficient of task completion, w 3 Represents the weight coefficient of energy consumption;

[0161] Here, safety, task completion, and energy consumption are key indicators of the drone’s performance after executing an action, and are used to construct the reward function;

[0162] Based on the obtained comprehensive reward value, the Bellman equation is used to update the expected cumulative reward of the drone in the future. The expression is:

[0163] B=R+γ×max(B 1 );

[0164] Among them, B represents the expected cumulative reward that the drone will receive in the future, max(B 1) represents the maximum value of the expected cumulative reward that the drone will receive in the future, γ represents the discount rate of the expected cumulative reward, a larger γ value (such as 0.9 or 0.99) means more emphasis on future rewards, while a smaller γ value (such as 0.5 or 0.7) means more emphasis on current immediate rewards;

[0165] Combined with the maximum expected cumulative reward, determine the obstacle avoidance action with the best long-term performance;

[0166] Obstacle avoidance actions include changing the heading angle of the drone to bypass obstacles in front;

[0167] Raise or lower the drone's altitude to avoid obstacles above or below;

[0168] Adjust the speed of the drone to avoid collisions with fast-approaching obstacles;

[0169] If the drone cannot avoid an obstacle in time, it can choose to hover or brake urgently and wait for the obstacle to be removed before continuing to fly;

[0170] Once the optimal obstacle avoidance action is selected, the action is transmitted to the drone's control system to actually perform the obstacle avoidance operation. The drone's control system will adjust the working status of its motors, propellers and other components according to the instructions to ensure that the drone avoids obstacles according to the predetermined action.

[0171] S5. Generate a flight path using a path planning algorithm according to the optimal obstacle avoidance action.

[0172] The following steps are included:

[0173] Based on the optimal obstacle avoidance action, combined with the density of static obstacles, the behavior pattern of dynamic obstacles and the specific flight mission requirements, a safe action set covering feasible actions and a safe distance for the drone to fly are generated;

[0174] The safe distance for drone flight refers to the minimum distance that should be maintained between the drone and obstacles. The specific safe distance needs to be determined based on the maximum speed and acceleration of the drone and the minimum turning radius.

[0175] Generally speaking, in low-altitude flight missions, the safety distance is usually set to 1 to 5 meters. Within this range, you can flexibly respond to complex environments while ensuring safety. If flying in densely populated areas, the safety distance can be set to more than 10 meters.

[0176] Define the UAV's dynamic constraints based on its physical limitations and motion characteristics;

[0177] Physical limitations include control error (the difference between the actual state of the drone and the desired state), control latency (the time difference between the sensor obtaining data and the control system issuing a command), and control frequency (the time difference between the sensor obtaining data and the control system issuing a command);

[0178] Movement characteristics include maximum climb rate, maximum thrust, endurance, etc.;

[0179] The maximum climb rate of a drone is determined by the balance between its vertical thrust and gravity. When the vertical thrust is greater than gravity, the drone can climb upwards, otherwise, the drone will descend.

[0180] The maximum thrust of a drone is determined by the power output of its motors and propellers. The greater the thrust, the higher the acceleration and climb rate of the drone;

[0181] Construct a three-dimensional environment grid based on the current position of the UAV, the target point, the environmental feature vector, and the dynamic constraints of the UAV, put the UAV into the three-dimensional environment network, and output the current position coordinates of the UAV;

[0182] A 3D environment mesh is a structured data representation method used to represent and model a 3D space. It divides the 3D space into a series of regular or irregular cells (usually called "voxels"), each of which can store information about the area, such as whether there are obstacles, the type of object, height, density, etc.

[0183] Initialize the starting flight node and target node in the three-dimensional environment grid, and define the open list and the closed list;

[0184] The open list is the set of nodes to be searched, and the closed list is the set of nodes that have been searched;

[0185] The heuristic cost from the start node to the target node is calculated using the Euclidean distance, expressed as:

[0186]

[0187] Among them, h(n) represents the heuristic cost from the starting node to the target point, n represents the starting node, g represents the target node, and x n Indicates the coordinate of the starting node on the x-axis, that is, the horizontal (east-west) coordinate, y n Indicates the coordinate of the starting node on the y-axis, that is, the vertical (north-south) coordinate, z n Indicates the coordinate of the starting node on the z-axis, that is, the height direction (up and down direction), x g Indicates the coordinate of the target node on the x-axis, y g Represents the coordinate of the target node on the y-axis, z gIndicates the coordinate of the target node on the z-axis;

[0188] The reason why we choose Euclidean distance to calculate the heuristic cost here is that it is one of the most commonly used heuristic cost functions, which is simple and computationally efficient. By assuming that the drone can fly directly to the target point in three-dimensional space, although the actual flight may be affected by obstacles and other constraints, this estimation can help the algorithm quickly narrow the search range and find a path close to the optimal one;

[0189] Based on the accumulation of the turning radius and speed limit in the UAV's dynamic constraints, the dynamic constraint cost is obtained, and the expression is:

[0190]

[0191] Where c(j) represents the dynamic constraint cost, v n represents the speed of the drone at the starting node n, a n represents the acceleration of the drone at node n, r n represents the turning radius of the drone at node n, v max It indicates the maximum permissible speed of the drone, that is, the maximum speed that the drone can reach under normal working conditions. max It represents the maximum permissible acceleration of the UAV, that is, the maximum acceleration that the UAV can withstand under normal working conditions, r min It represents the minimum allowable turning radius of the UAV, that is, the minimum turning radius that the UAV can perform under normal working conditions. α represents the weight coefficient of the speed term, which is used to adjust the influence of the speed on the total cost. β represents the weight coefficient of the acceleration term, which is used to adjust the influence of the acceleration on the total cost. δ represents the weight coefficient of the turning radius term, which is used to adjust the influence of the turning radius on the total cost.

[0192] Taking the environmental feature vector as the benchmark, the actual cost from the starting point to the start node in the three-dimensional environmental grid is obtained, and the expression is:

[0193]

[0194] r(i)=r(i`)+N

[0195] Among them, N represents the movement cost from the previous node to the starting node n, that is, the cost required for the drone to fly from the previous node to the current node, K represents the distance from the current node to the adjacent node, and v avg represents the average flight speed of the drone, r(i`) represents the cumulative cost from the starting point to the previous node of the starting node;

[0196] According to the heuristic cost from the start node to the target node, the dynamic constraint cost and the actual cost from the starting point to the start node, the total cost function of all nodes is obtained, which is expressed as:

[0197] f(g)=r(i)+h(n)+c(j);

[0198] Among them, f(g) represents the total cost function of all nodes, r(i) represents the actual cost from the starting point to the starting node in the three-dimensional environment grid, and c(j) represents the dynamic constraint cost;

[0199] Add the current position of the drone to the open list, and select the starting node with the smallest total cost function from the open list;

[0200] Select the A* algorithm to search the flight path. When the smallest starting node is equal to the target point, the search is successful and the flight path is output;

[0201] When the smallest starting node is not equal to the target point, the total cost function of the neighboring nodes is updated and added to the open list, the smallest starting node is moved to the closed list, and the operation is repeated until the flight path is found.

[0202] S6. During the flight, monitor the status of the drone and environmental changes in real time to perform collision detection and adjust the obstacle avoidance strategy.

[0203] The following steps are included:

[0204] The Kalman filter is a recursive algorithm that is widely used in predictive state estimation problems. Its main purpose is to predict the true state of the system and minimize the estimation error by fusing data from multiple sensors. The Kalman filter is used to predict the position of an obstacle in the future. The expression is:

[0205]

[0206] Among them, O represents the predicted position of the obstacle in the future, G represents the obstacle position at the current time t, V represents the obstacle speed at the current time t, Δt represents the predicted time step from the current time t to the future, usually in seconds. The control system of the drone needs a certain amount of time to perform obstacle avoidance actions (such as adjusting direction, changing altitude, etc.). Δt should be greater than or equal to the minimum reaction time of the drone to ensure that the system has enough time to make adjustments. For example, if the minimum reaction time of the drone is 0.2 seconds, then Δt should be set to at least 0.2 seconds. P represents the acceleration of the obstacle, which means the additional displacement caused by acceleration within Δt;

[0207] The predicted obstacle position information in the future is placed in the three-dimensional environment grid, and the straight-line distance between the current position of the drone and the obstacle position in the future is calculated. The expression is:

[0208]

[0209] Where T represents the straight-line distance between the current position of the drone and the position of the obstacle in the future, u represents the drone, m represents the obstacle, and x represents the distance between the current position of the drone and the obstacle in the future. u Indicates the coordinate of the drone on the x-axis, y u Indicates the coordinate of the drone on the y-axis, z u Indicates the coordinate of the drone on the z-axis, x m Represents the predicted position coordinates of the obstacle on the x-axis, y m Represents the predicted position coordinate of the obstacle on the y-axis, z m Indicates the predicted position coordinates of the obstacle on the z-axis;

[0210] When the straight-line distance between the current position of the drone and the position of the obstacle in the future is greater than the safe distance for the drone to fly, no collision occurs and the drone operates normally;

[0211] When the straight-line distance between the current position of the drone and the position of the obstacle in the future is less than or equal to the safe distance for the drone to fly, there is a possibility of collision. The A* algorithm is used again to search and plan a new flight path, which is sent to the flight controller to monitor environmental changes in real time. The flight direction, speed and altitude of the drone are adjusted according to actual needs. Repeat the operation to further adjust the flight path.

[0212] If a collision is unavoidable, a comprehensive emergency response mechanism is required to ensure the safety of the drone. Common emergency response measures include:

[0213] Lateral avoidance: The drone can quickly change its flight direction in a short period of time to avoid obstacles. This can be achieved by adjusting the attitude angle of the drone (such as roll, pitch, and yaw);

[0214] Vertical avoidance: If the horizontal avoidance space is limited, the drone can choose to rise or descend vertically to avoid obstacles;

[0215] Return to the take-off point and find a safe landing point: If the drone cannot continue to perform the current mission, it can choose to return to the take-off point to ensure that the drone can land safely. If returning to the take-off point is not feasible, the lidar sensor can be used to find a nearby flat ground as a temporary landing point to ensure that the drone can land safely.

[0216] Activate protective devices: Some drones are equipped with protective nets or other protective devices that can be deployed before a collision to reduce the damage to the drone;

[0217] Record accident data: Use the black box to record all sensor data (such as position, speed, acceleration, etc.) at the time of the accident, so as to analyze the cause of the accident and improve the performance of the drone;

[0218] To further explain, the collision detection method used here is based on two perspectives: image space and geometric space;

[0219] According to the results of collision detection and the complexity of the current environment, the parameters in the path planning algorithm are adjusted, including: when flying in densely populated areas or high-risk areas, by increasing the weight of safety, the drone will pay more attention to obstacle avoidance and maintaining a safe distance, even if this means sacrificing a certain degree of mission completion or increasing energy consumption. For example, when conducting aerial photography or logistics delivery in the city, the drone needs to avoid buildings, pedestrians and other vehicles. At this time, the weight of safety should be increased;

[0220] When the flight mission is time-sensitive or has strict time limits, the drone needs to reach the target location and complete the mission as quickly as possible. By increasing the weight of mission completion and taking a more aggressive flight path or speed, it can even tolerate higher energy consumption or slightly reduce safety requirements to a certain extent (but still ensure basic safety). For example, in a fire rescue mission, the drone needs to quickly find trapped people and provide assistance. At this time, the weight of mission completion should be increased to ensure that the drone can reach the target location as soon as possible.

[0221] When the mission requires long flight time or the drone’s battery capacity is limited, the drone extends its flight time. For example, in agricultural monitoring missions, the drone needs to conduct long inspections over large tracts of farmland. At this time, the weight of energy consumption should be high to ensure that the drone can complete the entire mission on a single charge.

[0222] At the same time, transfer learning technology is used to transfer knowledge that has been learned in similar environments to new environments. For example, if a drone learns how to avoid pedestrians and vehicles in city A, this knowledge can be transferred to the environment of city B.

[0223] This embodiment also provides a real-time scene collision detection system for low-altitude UAVs, including:

[0224] The data acquisition module uses a variety of sensors to collect real-time data about the surrounding environment of low-altitude drones;

[0225] The model generation module integrates the collected data to generate a unified environmental model;

[0226] The object classification module classifies objects in the scene based on the fused environment model;

[0227] The obstacle avoidance selection module selects the best obstacle avoidance action based on the classification results;

[0228] The path generation module generates the flight path using the path planning algorithm according to the optimal obstacle avoidance action;

[0229] Adjust the strategy module to monitor the status of the drone and environmental changes in real time during flight to perform collision detection and adjust the obstacle avoidance strategy.

[0230] This embodiment also provides a computer device, which is suitable for the real-time scene collision detection method for low-altitude unmanned aerial vehicles, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the real-time scene collision detection method for low-altitude unmanned aerial vehicles proposed in the above embodiment.

[0231] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0232] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the real-time scene collision detection method for a low-altitude unmanned aerial vehicle proposed in the above embodiment is implemented; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, referred to as EPROM), programmable read-only memory (Programmable Red-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0233] In summary, the present invention generates an environmental feature vector by combining semantic maps and target tracking results, sets a reward function based on the safety, mission completion and energy consumption of the drone, updates the expected cumulative reward of the drone in the future through the Bellman equation, and selects the best obstacle avoidance action. This decision-making method based on comprehensive reward value not only improves the safety of the drone, but also ensures the smooth completion of the flight mission, and extends the drone's flight time by optimizing energy consumption. In addition, the drone continuously learns and optimizes the obstacle avoidance strategy based on historical flight data, further improving the ability and efficiency of autonomous flight.

[0234] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A real-time scene collision detection method for low-altitude UAVs, characterized in that: include, Use a variety of sensors to collect real-time data about the surrounding environment of low-altitude drones; Fuse the collected data to generate a unified environmental model; Classify objects in the scene based on a unified environment model; Select the best obstacle avoidance action based on the classification results; A path planning algorithm is used to generate a flight path based on the best obstacle avoidance action; During the flight, the drone's status and environmental changes are monitored in real time to detect collisions and adjust obstacle avoidance strategies.

2. The real-time scene collision detection method for low-altitude unmanned aerial vehicles according to claim 1, characterized in that: Use a variety of sensors to collect real-time data about the surrounding environment of low-altitude drones, including the following steps: Collect historical flight data of drones in real time; The multiple sensors are laser radar, millimeter wave radar, camera, ultrasonic wave and GPS / IMU sensor; Use a laser radar sensor to emit a laser beam to the surrounding area, extract the three-dimensional space coordinates according to the angle of the laser beam, and output point cloud data; Use a millimeter-wave radar sensor to modulate the frequency of radar transmission and output radar data of the distance, speed and reflection intensity of the target point; Using a camera sensor to capture an image through a lens and output image data; Use ultrasonic sensors to transmit ultrasonic signals and receive echo signals, and output the distance data between the obstacle and the target point; Use GPS / IMU sensors to output position information data and attitude angle data respectively.

3. The real-time scene collision detection method for low-altitude unmanned aerial vehicles according to claim 2, characterized in that: The collected data is integrated to generate a unified environment model, including the following steps: Use statistical filtering algorithm to eliminate noise in point cloud data and perform sampling; Use the constant false alarm rate detection method to extract the distance, speed and angle information of the target point in the radar data and generate a target point set; De-noise and contrast enhance the image data and convert it into digital signals; The distance data between obstacles and target points that exceed the effective detection distance are eliminated through the threshold value; After processing, the boundaries, shape, and size of obstacles are extracted from the LiDAR sensor; Extract the speed and direction of motion of dynamic obstacles from the millimeter wave radar sensor; Extracting semantic and depth information from camera sensors; Extracting distance information of close obstacles from ultrasonic sensors; The extracted data from LiDAR, millimeter-wave radar, camera, and ultrasound are integrated to form the state vector of the obstacle. Based on the state vectors of obstacles, a unified environment model is generated.

4. The real-time scene collision detection method for low-altitude unmanned aerial vehicles according to claim 3, characterized in that: Based on the unified environment model, objects in the scene are classified, including the following steps: Select the convolutional neural network learning model and ReLU activation function; The state vector of the obstacle is input into the activation function, and the features of the obstacle are extracted after linear transformation; The obstacle features are linearly transformed again through the convolutional neural network learning model to obtain the classification probability distribution; According to the classification probability distribution, all pixels of the object image are processed one by one, and the classification probability of each pixel is calculated. The expression is: Among them, C represents the classification probability of each pixel, Z represents the normalization constant, σ represents the normalization function that converts the category score into a probability distribution, W represents the weight matrix, φ(S) represents the new feature vector after being processed by the activation function, and b represents the bias term; Based on the classification probability of each pixel, the maximum probability category of the object pixel is calculated and the category threshold of the object image pixel is set; When the maximum probability of an object pixel is greater than or equal to the category threshold of the object image pixel, the object is marked as dynamic; When the maximum probability of an object pixel is less than the category threshold of the object image pixel, the object is marked as static.

5. The real-time scene collision detection method for low-altitude unmanned aerial vehicles according to claim 4, characterized in that: Selecting the best obstacle avoidance action based on the classification results includes the following steps: Use deep convolutional neural networks to perform semantic segmentation on images captured by the camera and generate semantic maps; Based on the semantic map, the DeepSORT target tracking algorithm is used in combination with image data to extract the position and speed of dynamic obstacles and output the state vector of the dynamic obstacles. Combine the semantic map and target tracking results to generate an environment feature vector; According to the safety, mission completion and energy consumption of the drone during flight, the reward function is set and the comprehensive reward value is calculated. The expression is: R = w1×D+w2×E+w3×F; Among them, R represents the comprehensive reward value obtained by the drone when performing the corresponding action in flight, D represents the safety reward of the drone in flight, E represents the task completion reward of the drone in flight, F represents the energy consumption reward of the drone in flight, w1 represents the weight coefficient of safety, w2 represents the weight coefficient of task completion, and w3 represents the weight coefficient of energy consumption; Based on the obtained comprehensive reward value, the Bellman equation is used to update the expected cumulative reward of the drone in the future and combined with the maximum value of the expected cumulative reward to determine the obstacle avoidance action with the best long-term performance.

6. The real-time scene collision detection method for low-altitude unmanned aerial vehicles according to claim 5, characterized in that: According to the best obstacle avoidance action, the flight path is generated using the path planning algorithm. The following steps are included: Based on the optimal obstacle avoidance action, combined with the density of static obstacles, the behavior pattern of dynamic obstacles and the specific flight mission requirements, a safe action set covering feasible actions and a safe distance for the drone to fly are generated; Define the UAV's dynamic constraints based on its physical limitations and motion characteristics; Construct a three-dimensional environment grid based on the current position of the UAV, the target point, the environmental feature vector, and the dynamic constraints of the UAV, put the UAV into the three-dimensional environment network, and output the current position coordinates of the UAV; Initialize the starting flight node and target node in the three-dimensional environment grid, and define the open list and the closed list; The open list is a set of nodes to be searched, and the closed list is a set of nodes that have been searched; The heuristic cost from the start node to the target node is calculated using the Euclidean distance, expressed as: Among them, h(n) represents the heuristic cost from the starting node to the target point, n represents the starting node, g represents the target node, and x n Indicates the coordinate of the starting node on the x-axis, y n Indicates the coordinate of the starting node on the y-axis, z n Indicates the coordinate of the starting node on the z-axis, x g Indicates the coordinate of the target node on the x-axis, y g Indicates the coordinate of the target node on the y-axis, z g Indicates the coordinate of the target node on the z-axis; The dynamic constraint cost is obtained by accumulating the turning radius and speed limit in the dynamic constraint of the UAV; Taking the environmental feature vector as a benchmark, the actual cost from the starting point to the start node in the three-dimensional environmental grid is obtained; According to the heuristic cost from the start node to the target node, the dynamic constraint cost and the actual cost from the starting point to the start node, the total cost function of all nodes is obtained, which is expressed as: f(g)=r(i)+h(n)+c(j); Among them, f(g) represents the total cost function of all nodes, r(i) represents the actual cost from the starting point to the starting node in the three-dimensional environment grid, and c(j) represents the dynamic constraint cost; Add the current position of the drone to the open list, and select the starting node with the smallest total cost function from the open list; Select the A* algorithm to search the flight path. When the smallest starting node is equal to the target point, the search is successful and the flight path is output; When the smallest starting node is not equal to the target point, the total cost function of the neighboring nodes is updated and added to the open list, the smallest starting node is moved to the closed list, and the operation is repeated until the flight path is found.

7. The real-time scene collision detection method for low-altitude UAV according to claim 6, characterized in that: During the flight, the status of the drone and environmental changes are monitored in real time to detect collisions and adjust the obstacle avoidance strategy, including the following steps: Use the Kalman filter to predict the position of obstacles in the future. The expression is: Among them, O represents the predicted position of the obstacle in the future, G represents the obstacle position at the current time t, V represents the obstacle speed at the current time t, Δt represents the predicted time step from the current time t to the future, and P represents the acceleration of the obstacle; The predicted obstacle position information in the future is placed in the three-dimensional environment grid, and the straight-line distance between the current position of the drone and the obstacle position in the future is calculated. The expression is: Where T represents the straight-line distance between the current position of the drone and the position of the obstacle in the future, u represents the drone, m represents the obstacle, and x represents the distance between the current position of the drone and the obstacle in the future. u Indicates the coordinate of the drone on the x-axis, y u Indicates the coordinate of the drone on the y-axis, z u Indicates the coordinate of the drone on the z-axis, x m Represents the predicted position coordinates of the obstacle on the x-axis, y m Represents the predicted position coordinate of the obstacle on the y-axis, z m Indicates the predicted position coordinates of the obstacle on the z-axis; When the straight-line distance between the current position of the drone and the position of the obstacle in the future is greater than the safe distance for the drone to fly, no collision occurs and the drone operates normally; When the straight-line distance between the current position of the drone and the position of the obstacle in the future is less than or equal to the safe distance for the drone to fly, there is a possibility of collision. The A* algorithm is used again to search and plan a new flight path, which is sent to the flight controller to monitor environmental changes in real time. The flight direction, speed and altitude of the drone are adjusted according to actual needs. Repeat the operation to further adjust the flight path. Adjust the obstacle avoidance strategy according to the results of the collision detection.

8. A real-time scene collision detection system for low-altitude UAVs, based on the real-time scene collision detection method for low-altitude UAVs according to any one of claims 1 to 7, characterized in that: include, The data acquisition module uses a variety of sensors to collect real-time data about the surrounding environment of low-altitude drones; The model generation module integrates the collected data to generate a unified environmental model; The object classification module classifies objects in the scene based on the fused environment model; The obstacle avoidance selection module selects the best obstacle avoidance action based on the classification results; The path generation module generates the flight path using the path planning algorithm according to the optimal obstacle avoidance action; Adjust the strategy module to monitor the status of the drone and environmental changes in real time during flight to perform collision detection and adjust the obstacle avoidance strategy.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the real-time scene collision detection method for low-altitude unmanned aerial vehicles described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the real-time scene collision detection method for low-altitude unmanned aerial vehicles described in any one of claims 1 to 7 are implemented.

Images