Intelligent obstacle avoidance system and method for low-altitude aircraft

Abstract

The invention discloses an intelligent obstacle avoidance system and method for a low-altitude aircraft. Efficient obstacle avoidance is realized by fusing an improved attention mechanism and a binocular vision sensing multi-dimensional analysis model. In the system, a binocular vision perception unit collects double-path images and establishes motion association, a feature enhancement attention unit performs dynamic weighting according to image features, and a multi-dimensional analysis processing unit constructs a three-dimensional feature space to accurately recognize obstacles. The path planning decision-making unit combines the aircraft state and a safety threshold value, an optimal obstacle avoidance path is screened through heuristic search and cost evaluation, and the motion control instruction generation unit converts the path into a posture and speed control instruction to drive obstacle avoidance. According to the invention, the defects of the traditional technology in obstacle identification and path planning are overcome, dynamic or hidden obstacles can be accurately identified in a complex environment, the obstacle avoidance path is efficiently planned, and the operation safety and reliability of the low-altitude aircraft are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of low-altitude aircraft obstacle avoidance, and in particular to an intelligent obstacle avoidance system and method for low-altitude aircraft. BACKGROUND

[0002] With the wide application of low-altitude aircraft such as unmanned aerial vehicles and low-altitude manned aircraft in fields such as logistics transportation, environmental monitoring, and emergency rescue, the demand for safe obstacle avoidance of low-altitude aircraft in complex low-altitude environments is increasingly urgent. Low-altitude environments have the characteristics of diverse types of obstacles, complex distribution, and dynamic changes, such as buildings and power lines in cities, trees and rocks in mountainous areas, and moving birds and other aircraft, which pose high requirements on the obstacle avoidance capability of low-altitude aircraft.

[0003] However, the existing low-altitude aircraft obstacle avoidance technology has many shortcomings. On the one hand, there are limitations in obstacle recognition capability. Traditional obstacle avoidance systems mostly rely on a single sensor or simple multi-sensor fusion, which is difficult to comprehensively and accurately perceive environmental information. For example, although laser radar can quickly obtain distance information, it has weak recognition ability for small, transparent, or texture-similar obstacles in complex lighting and adverse weather conditions; and a simple visual sensor is easily affected by factors such as lighting changes and occlusion, resulting in missed or false detection of obstacles, and cannot provide reliable environmental perception data for the aircraft.

[0004] On the other hand, in terms of obstacle avoidance decision and path planning, the existing technology is difficult to adapt to complex and variable low-altitude scenes. Traditional path planning algorithms, such as A* algorithm and Dijkstra algorithm, have high computational complexity and poor real-time performance, which is difficult to meet the needs of rapid decision-making of the aircraft. At the same time, when dealing with dynamic obstacles and narrow and complex spaces, these algorithms often generate suboptimal paths, which may result in long paths, excessive turns, or even local optimal solutions, making the aircraft have collision risks in the obstacle avoidance process and unable to efficiently and safely complete the obstacle avoidance task. SUMMARY

[0005] In order to overcome the shortcomings and deficiencies of the prior art, the present application provides an intelligent obstacle avoidance system and method for low-altitude aircraft.

[0006] The technical solution adopted by the present application is an intelligent obstacle avoidance system for low-altitude aircraft, comprising:

[0007] A binocular vision perception unit based on spatiotemporal feature enhancement is used to synchronously collect dual-channel image information of the low-altitude environment in an interleaved pixel sampling mode, and establish pixel motion correlation between image frames through optical flow estimation to form multi-channel image tensor data;

[0008] The adaptive weight distribution feature enhancement attention unit is used to dynamically weight the features of each channel and spatial position of the multi-channel image tensor data according to the texture complexity, edge gradient intensity and target contour feature of the image region through the multi-level attention calculation module, to generate an enhanced feature tensor;

[0009] The multi-scale feature fusion multi-dimensional analysis processing unit is used to perform multi-scale convolution operation on the enhanced feature tensor, extract semantic and geometric features at different resolutions, and construct a three-dimensional feature space containing global semantics and local details through cross-scale feature splicing and fusion operation.

[0010] The obstacle recognition unit based on dynamic threshold determination is used to identify the obstacles in the environment and obtain the position coordinates, size information and confidence score by calculating the feature similarity and spatial distance threshold in the three-dimensional feature space using the preset obstacle feature template library and the sliding window search strategy.

[0011] The obstacle avoidance path planning decision unit is used to generate multiple candidate obstacle avoidance paths in the flight allowed area using a heuristic search algorithm according to the information output by the obstacle recognition unit, combined with the current position, velocity vector of the aircraft and the preset safety distance threshold.

[0012] The motion control instruction generation unit is used to convert the optimal obstacle avoidance path selected by the obstacle avoidance path planning decision unit into attitude adjustment parameters and speed control instructions of the aircraft, and calculate the motor speed regulation amount and rudder deflection angle through attitude solving algorithm and dynamics model calculation.

[0013] Further, in the adaptive weight distribution feature enhancement attention unit, the improved attention mechanism adopts the following model formula: Where A ij represents the attention weight of the i-th position to the j-th position; f i , f j , f k are the feature vectors of the i-th, j-th and k-th positions in the feature tensor; W1 and W2 are learnable weight matrices; b is a bias vector; σ is an activation function; and N is the total number of positions in the feature tensor.

[0014] Further, in the multi-scale feature fusion multi-dimensional analysis processing unit, the feature fusion formula based on the binocular vision sensing multi-dimensional analysis model is as follows: F fusion = α1F s1 + α2F s2 + α3F s3 Where F fusion is the fused feature tensor; F s1 , F s2 , F s3Respectively, the feature tensor extracted under different scales; α1, α2, α3 are adaptive fusion weights, which are calculated by the following formula: Wherein, m=1, 2, 3; l=1, 2, 3; Indicate the gradient norm of the nth feature map under the s scale; M is the number of feature maps.

[0015] Further, in the obstacle recognition unit based on dynamic threshold judgment, the improved obstacle recognition similarity calculation model formula is: Wherein, S ij Indicates the similarity of the ith to be identified region and the jth obstacle feature template; f i p , Respectively, the feature value of the ith to be identified region and the ith obstacle feature template in the pth feature dimension; P is the total number of feature dimensions.

[0016] Further, in the obstacle avoidance path planning decision unit, the path cost evaluation function formula is: C path =β1D obstacle +β2L path +β3A change , wherein, C path is the total path cost; D obstacle is the minimum distance between the path and the obstacle, which is calculated by the obstacle position information obtained by the binocular vision sensing multi-dimensional analysis model; L path is the path length; A change is the direction change amount of the path; β1, β2, β3 are weight coefficients, which are dynamically adjusted after evaluating the importance of different factors by the improved attention mechanism.

[0017] Further, in the binocular vision perception unit based on spatiotemporal feature enhancement, the optical flow estimation improved model formula is: Wherein, v x，y is the optical flow vector of pixel point (x, y); I t , I t+1 are the images at time t and t+1 respectively; m, n are the neighborhood pixel coordinate offset; M, N are the neighborhood window size.

[0018] Further, the feature enhancement attention unit with adaptive weight distribution further comprises a secondary attention optimization module, and the model formula is: Wherein, is the optimized attention weight; A ij is the attention weight calculated initially; τ is a temperature parameter for adjusting the distribution of attention weight; k is the kth accumulation, and N is the total number of accumulations.

[0019] Further, in the multi-scale feature fusion multi-dimensional analysis processing unit, a cross-channel attention mechanism is introduced, and the model formula is: Wherein, is the feature tensor after channel optimization; F c is the feature tensor of the original cth channel; F d is the feature tensor of the dth channel; D is the total number of channels; γ c is the channel reservation coefficient; ω cd is the inter-channel attention weight, which is calculated by the improved attention mechanism combined with the correlation of different channel features in the binocular vision sensing multi-dimensional analysis model.

[0020] Further, in the obstacle avoidance path planning decision unit, a hierarchical search strategy is adopted, and the search range adjustment formula is: Wherein, R search is the adjusted search range; R base is the basic search range; D nearest is the distance between the aircraft and the nearest obstacle, which is determined by the binocular vision sensing multi-dimensional analysis model; D safe is the preset safety distance; δ is the proportional coefficient.

[0021] The obstacle avoidance method of the intelligent obstacle avoidance system comprises the following steps:

[0022] Step S1, using the binocular vision perception unit based on space-time domain feature enhancement to synchronously collect the dual-channel image information of the low-altitude environment in the staggered pixel sampling mode, and establishing the pixel motion correlation between the image frames through optical flow estimation to form multi-channel image tensor data;

[0023] Step S2, inputting the multi-channel image tensor data into the feature enhancement attention unit with adaptive weight distribution, and through the multi-level attention calculation module, dynamically weighting the features of each channel and spatial position according to the texture complexity, edge gradient intensity and target contour feature of the image area, to generate the enhanced feature tensor;

[0024] Step S3, transmitting the enhanced feature tensor to the multi-scale feature fusion multi-dimensional analysis processing unit, performing multi-scale convolution operation, extracting semantic and geometric features at different resolutions, and constructing a three-dimensional feature space containing global semantics and local details through cross-scale feature splicing and fusion operation;

[0025] Step S4, sending the three-dimensional feature space data into the obstacle identification unit based on dynamic threshold determination, using the preset obstacle feature template library in the space, adopting the sliding window search strategy, calculating the feature similarity and spatial distance threshold, identifying the obstacles in the environment, and obtaining their position coordinates, size information and confidence score;

[0026] Step S5, the obstacle avoidance path planning decision unit generates multiple candidate obstacle avoidance paths in the flight allowed area according to the information output by the obstacle identification unit, in combination with the current position and speed vector of the aircraft and the preset safety distance threshold, by using a heuristic search algorithm, and screens out an optimal obstacle avoidance path through a path cost evaluation function;

[0027] Step S6, the motion control instruction generation unit converts the optimal obstacle avoidance path into attitude adjustment parameters and speed control instructions of the aircraft, calculates motor speed regulation and rudder deflection angle through an attitude solving algorithm and a dynamics model, and drives the aircraft to perform an obstacle avoidance action.

[0028] Beneficial effects: The intelligent obstacle avoidance system and method for low-altitude aircraft proposed in the application, in the aspect of obstacle identification, the binocular vision perception unit based on spatiotemporal feature enhancement synchronously collects double images by using staggered pixel sampling, establishes inter-frame motion correlation by using optical flow estimation, dynamically weights each channel feature according to image texture, edge and contour features by using a feature enhancement attention unit with adaptive weight distribution, enhances the feature expression of key areas, and makes up for the problem of insufficient identification of special obstacles by a single sensor under complex lighting and severe weather; the multi-dimensional analytical processing unit with multi-scale feature fusion further extracts different resolution semantic and geometric features, constructs a three-dimensional feature space, and significantly improves the accuracy and integrity of obstacle identification by using a dynamic threshold determination mechanism. In the aspects of obstacle avoidance decision and path planning, the obstacle avoidance path planning decision unit generates candidate paths by using a heuristic search according to the obstacle identification result, in combination with the real-time state of the aircraft and the preset safety threshold, screens out an optimal path by using a cost evaluation strategy that comprehensively considers obstacle avoidance safety, path length and direction change, and simultaneously improves the path planning efficiency by using a hierarchical search strategy that dynamically adjusts the search range according to the distance between the aircraft and the obstacle. The motion control instruction generation unit converts the optimal path into precise attitude and speed control instructions to drive the aircraft to safely avoid obstacles. The whole system realizes accurate identification and efficient obstacle avoidance of obstacles in a complex low-altitude environment, greatly improves the safety and reliability of low-altitude aircraft operation in a complex scene. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The system unit composition diagram of the application;

[0030] Figure 2 The method step flowchart of the application. DETAILED DESCRIPTION

[0031] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict, and the present application will be further described in detail below in combination with the drawings and specific embodiments.

[0032] As Figure 1As shown, the intelligent obstacle avoidance system of the low-altitude aircraft comprises:

[0033] The binocular vision perception unit based on spatiotemporal feature enhancement is used for synchronously collecting double-channel image information of a low-altitude environment in an interleaved pixel sampling mode, and establishing pixel motion correlation between image frames through optical flow estimation to form multi-channel image tensor data.

[0034] Specifically, as a basic module for collecting system environment information, the unit uses two high-resolution industrial-grade CMOS cameras for double-channel image collection. The cameras have a global shutter technology, a frame rate of 60 frames / s, a resolution of 2560x1920 pixels, a pixel size of 3.2 μm, and can clearly capture environmental details within a range of 50 meters and effectively identify obstacles with a diameter of not less than 15 centimeters. The parallelism of the optical axes of the two cameras is controlled within 0.02°, and the baseline distance is fixed at 350 mm, ensuring the accuracy of parallax calculation. Through the interleaved pixel sampling mode, combined with the hardware synchronization circuit, sub-millisecond synchronous collection is realized to obtain double-channel image data containing RGB color information. At the same time, the built-in optical flow estimation algorithm is based on a multi-layer pyramid structure, which can calculate the motion vector of the pixel points between adjacent frames, with a motion detection accuracy of 0.05 pixels / frame, and can identify dynamic targets with a speed as low as 0.3 m / s, finally generating multi-channel image tensor data containing color, depth and motion information, providing multi-dimensional raw information for subsequent processing.

[0035] The implementation process of the unit strictly follows the parameter calibration and data synchronization process. In the installation stage, the internal and external parameters of the cameras need to be calibrated, and the Zhang calibration method is used to obtain the accurate camera intrinsic matrix and distortion coefficient. The external parameter calibration ensures that the rotation and translation matrix errors of the two cameras are within the acceptable range. When working, the image acquisition module synchronously triggers the two cameras at a fixed frequency, and the collected data is transmitted to the processing unit through gigabit Ethernet. In the data preprocessing link, through operations such as noise removal and white balance correction, the data quality is guaranteed. The optical flow estimation module processes the continuous frame images in layers, starting from the low-resolution image to calculate the preliminary motion vector, and then gradually refining to the original resolution, effectively improving the calculation efficiency and accuracy, and finally outputting the processed multi-channel image tensor data to the subsequent unit for feature enhancement and analysis.

[0036] The feature enhancement attention unit with adaptive weight allocation is connected with the binocular vision perception unit based on spatiotemporal feature enhancement, and is used for dynamically weighting the features of each channel and spatial position of the multi-channel image tensor data through a multi-level attention calculation module according to the texture complexity, edge gradient intensity and target contour features of the image area, to generate an enhanced feature tensor.

[0037] Specifically, the unit focuses on feature enhancement of multi-channel image tensor data, and its core architecture includes two sub-modules of spatial attention and channel attention. The spatial attention module assigns weights based on the texture complexity and edge gradient intensity of the image region. The texture feature is calculated by the local binary pattern (LBP) algorithm, which quantizes the local texture information using an 8-neighborhood pattern. The edge is detected using the Sobel operator, and the gradient intensity is calculated using a 3x3 template. This can accurately identify the areas with rich texture and significant edges in the image and assign higher weights to highlight key information such as obstacle outlines. The channel attention module evaluates the importance of each feature channel based on their correlation. The Pearson correlation coefficient is used to measure the semantic association between channels, and the channels closely related to obstacle recognition are given more weight to suppress irrelevant information and achieve targeted enhancement of data features.

[0038] The implementation of the unit adopts a hierarchical iterative optimization strategy. In the initialization phase, the pre-trained model is used to assign initial weight parameters to each attention module. The weight matrix of the spatial attention module matches the size of the image feature map, and the weight vector of the channel attention module corresponds to the number of feature channels. During operation, the system updates the weights using the stochastic gradient descent algorithm based on the real-time input image tensor data. The step size is set to 0.01, the learning rate is 0.001, and the feature loss function is minimized through backpropagation to continuously optimize the weight distribution. After processing by the unit, the features related to obstacles in the original image tensor data are significantly enhanced, and irrelevant background information is effectively suppressed. The generated enhanced feature tensor provides more discriminative input data for subsequent obstacle recognition.

[0039] The multi-dimensional analysis processing unit of multi-scale feature fusion is connected with the feature enhancement attention unit with adaptive weight distribution, which is used to perform multi-scale convolution operation on the enhanced feature tensor to extract semantic and geometric features at different resolutions. Through cross-scale feature stitching and fusion operations, a three-dimensional feature space containing global semantics and local details is constructed.

[0040] Specifically, the unit is based on a multi-branch convolutional neural network architecture, realizing multi-scale feature extraction and fusion of the enhanced feature tensor. The unit contains three parallel convolution branches, each responsible for processing different scale features. The first branch uses a 3x3 convolution kernel with a step size of 1 and a padding of 1, focusing on extracting local detail features of the image, which can capture information such as obstacle surface texture and small structure; the second branch uses a 5x5 convolution kernel with a step size of 2 and a padding of 2, used to extract medium-scale semantic features, which can identify part of the structure of the obstacle; the third branch is equipped with a 7x7 convolution kernel with a step size of 4 and a padding of 3, mainly extracting global semantic features, used to judge the overall category and outline of the obstacle. The feature maps output by each branch are stabilized by the batch normalization layer, and the nonlinearity is introduced by the ReLU activation function to enhance the network expression ability. Finally, the channel splicing and weighted summation method is used to fuse different scale features, constructing a three-dimensional feature space containing global semantics and local details.

[0041] In the implementation process, the parameter configuration of each convolution branch is strictly optimized. The number of convolution kernels is set to 64, 128, and 256 respectively to balance the amount of calculation and feature extraction capability; the momentum parameter of the batch normalization layer is set to 0.9 to ensure the stability of data normalization; the fusion weight of weighted summation is obtained through training learning, and the initial value is randomly initialized with uniform distribution. To avoid overfitting, the Dropout strategy is used during network training, randomly discarding 20% of the neuron connections. The final generated three-dimensional feature space data not only retains the local detail features of the obstacle, but also integrates the global semantic information, providing rich feature dimensions for subsequent obstacle recognition and improving the system's recognition ability for different types and scales of obstacles.

[0042] The obstacle recognition unit based on dynamic threshold determination is connected with the multi-dimensional analysis processing unit of the multi-scale feature fusion, used to identify the obstacles in the environment by using the preset obstacle feature template library in the three-dimensional feature space, using the sliding window search strategy, calculating the feature similarity and spatial distance threshold, and obtaining the position coordinates, size information and confidence score of the obstacles;

[0043] Specifically, the unit relies on a pre-set obstacle feature template library to achieve accurate identification of obstacles in the environment in a three-dimensional feature space. The template library contains feature descriptions of different types of obstacles, such as spherical, cylindrical, planar structures, etc. Each template is composed of multiple feature vectors, covering shape, texture, spatial position, etc. The identification process uses a sliding window search strategy, with the window size determined by the resolution of the three-dimensional feature space. The minimum window coverage area corresponds to a spatial range of 0.2x0.2x0.2 cubic meters in the real scene. The unit calculates the similarity between the features of the area to be identified and the features in the template library, and combines the spatial distance threshold for judgment. The similarity calculation uses a combination of cosine distance and Manhattan distance to evaluate the direction and numerical difference of the feature vectors; the spatial distance threshold is dynamically adjusted according to the current speed of the aircraft, with the threshold increasing by 0.5 meters for every 1 meter / second increase in speed, ensuring timely response to close-range fast-moving obstacles.

[0044] When the unit is implemented, the three-dimensional feature space data is first divided into grids with a grid precision of 0.1 meters x 0.1 meters x 0.1 meters to improve search efficiency. The sliding window traverses each grid in turn, with each move step being 50% of the window size to ensure no missed detection. When the similarity of the area to be identified and the template exceeds the dynamically set threshold, and its spatial distance meets the safety requirements, it is determined as an obstacle, and its position coordinates (accuracy up to 0.1 meters), size information (error not more than 5%) and confidence score (value range 0-1) are recorded. To adapt to environmental changes, the system updates the template library every 5 seconds, merges similar templates through clustering algorithms, eliminates redundant information, and ensures the timeliness and accuracy of the template library, thereby achieving reliable identification of various obstacles in complex environments.

[0045] The obstacle avoidance path planning decision unit is connected to the obstacle identification unit based on dynamic threshold determination, and is used to generate multiple candidate obstacle avoidance paths in the flight allowed area according to the information output by the obstacle identification unit, combined with the current position, speed vector of the aircraft and the pre-set safety distance threshold, using a heuristic search algorithm.

[0046] Specifically, the unit generates a safe and efficient obstacle avoidance path based on the obstacle recognition result and the real-time state parameters of the aircraft. The unit receives information such as the current position coordinates of the aircraft (accuracy 0.1 meters), velocity vector (accuracy 0.01 meters / second), heading angle (accuracy 0.1°), and reference to the preset safety distance threshold (default value 3 meters, adjustable according to environmental complexity). A heuristic search algorithm is used to generate multiple candidate obstacle avoidance paths in the feasible region. The search process is based on octree space division, with a tree node resolution of 0.5 meters x 0.5 meters x 0.5 meters to ensure the fineness of path planning. The candidate path evaluation considers factors such as obstacle avoidance safety, path length, and direction change. Safety is measured by calculating the minimum distance between the path and the obstacle, path length is calculated using Euclidean distance, and direction change is evaluated by the cumulative angle of adjacent path segments. Through a preset cost evaluation strategy, each candidate path is scored, and the one with the lowest score is determined as the optimal obstacle avoidance path.

[0047] In the implementation process, the unit uses a hierarchical search strategy to improve computational efficiency. In the initial stage, a coarse-grained search is performed with a larger search step (1 meter) to quickly filter out the approximate feasible direction; then a fine search is performed in the key area with a smaller step (0.2 meters) to optimize the path details. The search range is dynamically adjusted according to the distance between the aircraft and the nearest obstacle. For every 1 meter reduction in distance, the search range is expanded by 20%, ensuring timely response to sudden obstacles. To avoid falling into a local optimal solution, a random perturbation mechanism is introduced into the algorithm, which adjusts the current optimal path locally with a probability of 5% in each iteration. The final optimal obstacle avoidance path contains a series of discrete path points, with a distance between adjacent points not exceeding 0.5 meters, providing accurate guidance for the motion control of the aircraft.

[0048] The motion control instruction generation unit is connected to the obstacle avoidance path planning decision unit and is used to convert the optimal obstacle avoidance path selected by the obstacle avoidance path planning decision unit into attitude adjustment parameters and speed control instructions for the aircraft. Through attitude solving algorithms and dynamic models, the motor speed adjustment amount and rudder deflection angle are calculated.

[0049] Specifically, the unit is responsible for converting the optimal obstacle avoidance path into control instructions executable by the aircraft to achieve precise obstacle avoidance actions. The unit receives discrete path point coordinate information of the optimal obstacle avoidance path, combines the current attitude parameters of the aircraft (pitch angle, roll angle, yaw angle, accuracy 0.1°) and the dynamics model, and calculates the aircraft attitude adjustment parameters through an attitude solving algorithm. The attitude adjustment adopts a proportional-integral-derivative (PID) control strategy, with proportional coefficient, integral coefficient and derivative coefficient set to 0.8, 0.2 and 0.1 respectively, to ensure the stability and rapidity of attitude adjustment. The speed control instruction is planned according to the distance between path points and the preset maximum speed (default value is 10 m / s), and a trapezoidal acceleration-deceleration algorithm is used to control the speed change, with acceleration and deceleration set to 2 m / s2 to avoid impact on the aircraft caused by sudden speed changes.

[0050] In the implementation process, the unit converts the attitude adjustment parameters and the speed control instruction into motor speed regulation and rudder deflection angle. For a quadcopter, attitude adjustment is achieved by changing the speed difference of the four motors, with a speed regulation accuracy of 1 revolution per second; the control accuracy of the rudder deflection angle is 0.5°, ensuring that the aircraft can fly smoothly according to the planned path. The instruction output uses pulse width modulation (PWM) signals, with a signal frequency of 50 Hz and a duty cycle regulation range of 5%-15% to adapt to the driving system of the aircraft. At the same time, the unit has a built-in safety protection mechanism that automatically triggers an emergency braking program when the control instruction exceeds the performance limit of the aircraft, ensuring the safety of the aircraft operation.

[0051] Preferably, in the feature enhancement attention unit with adaptive weight distribution, the improved attention mechanism adopts the following model formula: where A ij represents the attention weight of the i-th position to the j-th position; f i , f j , f k are the feature vectors of the i-th, j-th and k-th positions in the feature tensor; W1 and W2 are learnable weight matrices; b is a bias vector; σ is an activation function; and N is the total number of positions in the feature tensor. This formula realizes the enhancement of important feature regions by weighted sum of different position features, and provides more discriminative feature representation for obstacle identification by combining the texture and edge information of the image in the multi-dimensional analysis model of binocular vision sensing.

[0052] Specifically, the improved attention mechanism in the feature enhancement attention unit of the adaptive weight allocation enhances the important feature regions of the image by calculating the weights of different positions in the feature tensor. In the implementation process, the feature vector at each position in the feature tensor is taken as input, and a learnable weight matrix and a bias vector are used to perform linear transformation on it, and then the attention weight is calculated after processing by an activation function. This mechanism combines the image texture and edge information obtained by the multi-dimensional analysis model of binocular vision sensing to give higher weight to regions with complex texture and obvious edges, so that in the obstacle recognition, the features related to the obstacle can be highlighted and the background information interference can be suppressed, providing more discriminative feature representation for subsequent obstacle recognition and improving the accuracy of feature extraction.

[0053] Preferably, in the multi-dimensional analysis processing unit of the multi-scale feature fusion, the feature fusion formula based on the multi-dimensional analysis model of binocular vision sensing is F fusion = a1F s1 + a2F s2 + a3F s3 , where F fusion is the fused feature tensor; F s1 , F s2 , and F s3 are feature tensors extracted at different scales; a1, a2, and a3 are adaptive fusion weights, which are calculated by the following formula: where m = 1, 2, 3; l = 1, 2, 3; indicates the gradient norm of the nth feature map at the s-th scale; M is the number of feature maps; this formula dynamically adjusts the fusion weights according to the gradient information of different scale features, combines the improved attention mechanism to focus on important regions, effectively fuses the obstacle features at different scales, and improves the accuracy and completeness of obstacle recognition.

[0054] Specifically, in the multi-dimensional analysis processing unit of the multi-scale feature fusion, the feature fusion formula based on the multi-dimensional analysis model of binocular vision sensing realizes adaptive fusion of features at different scales. In actual application, different size convolution kernels are used to perform convolution operations on the feature tensor to extract feature tensors at different scales, such as small-scale convolution kernels to extract local detail features and large-scale convolution kernels to extract global semantic features. Then, the adaptive fusion weights are calculated according to the gradient norms of the feature maps at different scales. The greater the gradient norm, the richer the information contained in the feature at that scale, and the higher the corresponding fusion weight. Different scale features are fused by weighted summation, combined with the improved attention mechanism to focus on important regions, effectively integrating obstacle features at different scales, so that the fused feature tensor contains not only rich detail information but also global semantic information, improving the completeness and accuracy of obstacle recognition.

[0055] Preferably, in the obstacle recognition unit based on dynamic threshold determination, the improved obstacle recognition similarity calculation model formula is: Among them, S ij represents the similarity between the i-th area to be identified and the j-th obstacle feature template; f i p 、 are the eigenvalues ​​of the i-th area to be identified and the j-th obstacle feature template in the p-th feature dimension respectively; P is the total number of feature dimensions; combined with the three-dimensional spatial information obtained by the binocular visual sensing multi-dimensional analytical model, accurate identification of different types of obstacles is achieved by setting the dynamic similarity threshold and spatial distance threshold.

[0056] Specifically, the improved obstacle recognition similarity calculation model within the obstacle recognition unit, based on dynamic threshold determination, achieves accurate obstacle identification by calculating the similarity between the area to be identified and the obstacle feature template across multiple feature dimensions. During implementation, the feature vector of the area to be identified, obtained by the binocular vision sensing multidimensional analytical model, and the preset obstacle feature template vector are multiplied element-wise in each feature dimension and summed. The sum is then divided by the product of the module lengths of the two vectors to determine their similarity. Dynamic similarity and spatial distance thresholds are set in conjunction with the three-dimensional spatial information acquired through binocular vision. When the similarity between the area to be identified and the template exceeds the threshold and the spatial distance meets safety requirements, it is identified as an obstacle, enabling accurate identification of obstacles of different types and postures.

[0057] Preferably, in the obstacle avoidance path planning decision unit, the path cost evaluation function formula is: C path =β1D obstacle +β2L path +β3A change , where C path is the total path cost; D obstacle is the minimum distance between the path and the obstacle, which is calculated by the obstacle position information obtained by the binocular vision sensing multidimensional analytical model; L path is the path length; A change is the directional change of the path; β1, β2, and β3 are weight coefficients, which are dynamically adjusted after evaluating the importance of different factors through the improved attention mechanism; this formula comprehensively considers obstacle avoidance safety, path length, and directional change factors, and combines binocular visual sensing information with the improved attention mechanism to screen out the optimal obstacle avoidance path.

[0058] Specifically, the path cost evaluation function in the obstacle avoidance path planning decision unit comprehensively considers the distance between the path and the obstacle, the path length, and the path direction change to evaluate the pros and cons of the candidate obstacle avoidance path. In the actual path planning process, the position information of the obstacle obtained by the binocular vision sensing multi-dimensional analysis model is used to calculate the minimum distance between each candidate path and the obstacle; the path length is calculated according to the Euclidean distance of each path point on the path; and the direction change of the path is quantified by the included angle of adjacent path segments. The improved attention mechanism is used to evaluate the importance of different factors, and then dynamically adjust the weight coefficients corresponding to each factor. The total cost of each candidate path is calculated, and the path with the lowest total cost is selected as the optimal obstacle avoidance path, ensuring that the planned path achieves a good balance in safety, length, and direction change.

[0059] Preferably, in the binocular vision perception unit based on spatiotemporal feature enhancement, the optical flow estimation improved model formula is: wherein, v x，y is the optical flow vector of the pixel point (x, y); I t , I t+1 are the images at time t and t+1, respectively; m and n are the neighborhood pixel coordinate offsets; M and N are the neighborhood window sizes; the formula combines the improved attention mechanism to focus on the image calibration area, calculates the pixel changes of adjacent frames of images, accurately obtains the motion information of objects in the scene, and provides dynamic environment data for obstacle identification and path planning.

[0060] Specifically, the optical flow estimation improved model in the binocular vision perception unit based on spatiotemporal feature enhancement calculates the gray level changes of pixel points in the neighborhood range in adjacent frames of images to obtain the optical flow vector, thereby accurately obtaining the motion information of objects in the scene. In the implementation process, a neighborhood window of a certain size is set with each pixel point as the center, the gray level difference of the corresponding pixel points in the window is calculated, and the coordinate offset of the pixel points in the window is weighted and summed, and then divided by the square sum of the coordinate offset to obtain the optical flow vector of the pixel point. Combined with the improved attention mechanism to focus on the key areas of the image, higher weights are given to important areas when calculating the optical flow, so that the optical flow estimation result can better reflect the real motion situation of the objects in the key areas, and provide reliable dynamic environment data for obstacle identification and path planning.

[0061] Preferably, the feature enhancement attention unit with adaptive weight distribution further comprises a secondary attention optimization module, and the model formula thereof is: wherein, is the optimized attention weight; A ijis the attention weight obtained by the first calculation; τ is a temperature parameter, used to adjust the distribution of the attention weight; k is the kth accumulation, and N is the total number of accumulations; the formula combines the image feature information obtained by the binocular vision sensing multi-dimensional analysis model, and further highlights the calibration feature region through secondary optimization of the initial attention weight, thereby improving the extraction ability of the obstacle features in a complex environment.

[0062] Specifically, the secondary attention optimization module in the adaptive weight distribution feature enhancement attention unit optimizes the attention weight obtained by the first calculation again to further highlight the key feature region in the image. In actual operation, the attention weight obtained by the first calculation is taken as input, scaled by introducing a temperature parameter, and then normalized to obtain the optimized attention weight. The temperature parameter can be adjusted according to the actual scene, and the smaller the value, the more concentrated the distribution of the optimized attention weight, and the higher the attention degree to the key region. The module combines the image feature information obtained by the binocular vision sensing multi-dimensional analysis model, and through secondary optimization, the weight of the feature region related to the obstacle can be further enhanced in a complex environment, thereby effectively improving the extraction ability of the obstacle features in a complex background.

[0063] Preferably, in the multi-dimensional analysis processing unit of the multi-scale feature fusion, a cross-channel attention mechanism is introduced, and the model formula is: wherein, is the feature tensor after channel optimization; F c is the feature tensor of the original cth channel; F d is the feature tensor of the dth channel; D is the total number of channels; γ c is the channel reservation coefficient; ω cd is the inter-channel attention weight, which is obtained by improving the attention mechanism combined with the correlation of different channel features in the binocular vision sensing multi-dimensional analysis model; the formula realizes adaptive fusion of different channel features and enhances the expression ability of the multi-dimensional features of the obstacle.

[0064] Specifically, the cross-channel attention mechanism introduced in the multi-dimensional analysis processing unit of the multi-scale feature fusion realizes adaptive fusion of different channel features and enhances the expression ability of multi-dimensional features of obstacles. In the implementation process, for the original feature tensor of each channel, the correlation with all other channel feature tensors is calculated to obtain the inter-channel attention weight, and the correlation calculation can use methods such as Pearson correlation coefficient. Then, according to the channel retention coefficient and the inter-channel attention weight, the original channel feature is weighted and fused to obtain the optimized channel feature tensor. The improved attention mechanism is used to determine the correlation between channels and the channel retention coefficient, so that the channel features closely related to obstacle recognition are enhanced, and irrelevant channel features are suppressed, effectively integrating the multi-dimensional information such as obstacle texture and shape contained in different channels.

[0065] Preferably, in the obstacle avoidance path planning decision unit, a hierarchical search strategy is adopted, and the search range adjustment formula is: wherein, R search is the adjusted search range; R base is the basic search range; D nearest is the distance between the aircraft and the nearest obstacle, determined by the multi-dimensional analysis model of binocular vision sensing; D safe is the preset safety distance; and δ is the proportional coefficient. Combined with the improved attention mechanism for evaluating the importance of obstacles, the search range is dynamically adjusted according to the distance between the aircraft and the obstacles, improving the efficiency and safety of path planning.

[0066] Specifically, the hierarchical search strategy and its search range adjustment formula used in the obstacle avoidance path planning decision unit can dynamically adjust the search range according to the distance between the aircraft and the obstacles, improving the efficiency and safety of path planning. In actual path planning, the adjustment coefficient is calculated according to the distance between the aircraft and the nearest obstacle and the preset safety distance, and then the adjusted search range is obtained by combining the proportional coefficient and the basic search range. When the aircraft is close to the obstacle, the adjustment coefficient increases, and the search range expands accordingly, ensuring that potential obstacle avoidance space can be discovered in time; when the distance is far, the search range is reduced, reducing the calculation amount and improving the planning speed. Combined with the improved attention mechanism for evaluating the importance of obstacles, fine search is preferentially performed in the vicinity of key obstacles, so that the obstacle avoidance path is quickly and safely planned in complex environments.

[0067] As shown in Figure 2 , the obstacle avoidance method of the intelligent obstacle avoidance system comprises the following steps:

[0068] Step S1, using the binocular vision perception unit based on spatiotemporal feature enhancement to synchronously collect double-channel image information of the low-altitude environment in a staggered pixel sampling mode, and establishing pixel motion correlation between image frames through optical flow estimation to form multi-channel image tensor data;

[0069] Step S2, input the multi-channel image tensor data into the adaptive weight distribution feature enhancement attention unit, pass through the multi-level attention calculation module, and dynamically weight the features of each channel and spatial position according to the texture complexity, edge gradient intensity and target contour feature of the image region to generate an enhanced feature tensor;

[0070] Step S3, transmit the enhanced feature tensor to the multi-dimensional analysis processing unit of multi-scale feature fusion for multi-scale convolution operation to extract semantic and geometric features at different resolutions, and construct a three-dimensional feature space containing global semantics and local details through cross-scale feature splicing and fusion operation;

[0071] Step S4, send the three-dimensional feature space data into the obstacle recognition unit based on dynamic threshold judgment, use the preset obstacle feature template library in the space, adopt the sliding window search strategy, calculate the feature similarity and spatial distance threshold, identify the obstacles in the environment, and obtain their position coordinates, size information and confidence score;

[0072] Step S5, the obstacle avoidance path planning decision unit generates multiple candidate obstacle avoidance paths in the flight allowed area by using a heuristic search algorithm according to the information output by the obstacle recognition unit, combines the current position, velocity vector of the aircraft and the preset safety distance threshold, and selects the optimal obstacle avoidance path through a path cost evaluation function;

[0073] Step S6, the motion control instruction generation unit converts the optimal obstacle avoidance path into attitude adjustment parameters and speed control instructions of the aircraft, calculates the motor speed regulation amount and rudder deflection angle through attitude solving algorithm and dynamics model, and drives the aircraft to perform obstacle avoidance action.

[0074] The low-altitude aircraft intelligent obstacle avoidance system and method deeply fuse the improved attention mechanism and the multi-dimensional analysis model of binocular vision sensing, and construct a multi-level collaborative perception system to overcome the shortcomings of traditional systems in the field of obstacle recognition. The binocular vision sensing unit based on spatiotemporal feature enhancement adopts staggered pixel sampling and optical flow estimation technology to simultaneously collect double images and establish inter-frame motion correlation. Compared with a single sensor, it can more comprehensively capture environmental dynamic changes and detailed information. The feature enhancement attention unit with adaptive weight distribution dynamically weights the multi-channel image tensor data according to the image texture complexity, edge gradient intensity and target contour feature to enhance the key area feature expression. Combined with the multi-dimensional analysis processing unit of multi-scale feature fusion, three-dimensional feature space is constructed by extracting semantic and geometric features at different resolutions, and the dynamic threshold judgment mechanism is used to accurately identify small and transparent obstacles, or in complex lighting and harsh weather environments, effectively avoiding missed detection and false detection problems.

[0075] In terms of obstacle avoidance decision and path planning, the defects of traditional algorithms such as poor real-time performance and insufficient path optimization are effectively solved. The obstacle avoidance path planning decision unit is based on the obstacle recognition result, combined with the real-time state of the aircraft and the preset safety threshold, uses heuristic search to quickly generate candidate paths, and through the comprehensive consideration of the safety of obstacle avoidance, path length and the cost evaluation strategy of direction change, the optimal path is selected. At the same time, the hierarchical search strategy dynamically adjusts the search range according to the distance between the aircraft and the obstacle, which significantly improves the calculation efficiency. The motion control instruction generation unit accurately converts the optimal path into attitude adjustment parameters and speed control instructions, calculates the motor speed regulation and rudder deflection angle through the attitude solving algorithm and the dynamics model, ensures that the aircraft can flexibly and efficiently complete the obstacle avoidance action in complex and narrow space, avoids falling into a local optimal solution, and greatly enhances the adaptability and safety of the aircraft in a dynamic and complex environment.

[0076] In the description of the present application, it should be pointed out that, unless otherwise explicitly specified and limited, the terms "arrangement", "installation", "connection", "connection", "fixing" should be understood broadly, for example, it can be fixed connection, or detachable connection, or integral connection; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through intermediate medium, or the communication between two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood through specific circumstances.

[0077] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various equivalent changes, modifications, replacements and variations of the embodiments can be made without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalent scope.

Claims

1. The intelligent obstacle avoidance system for low-altitude aircraft is characterized by: include: A binocular visual perception unit based on spatiotemporal feature enhancement is used to synchronously collect dual-channel image information of low-altitude environments in an interleaved pixel sampling mode, and establish pixel motion correlation between image frames through optical flow estimation to form multi-channel image tensor data; The feature-enhanced attention unit with adaptive weight allocation is used to dynamically weight the features of each channel and spatial position of the multi-channel image tensor data according to the texture complexity, edge gradient strength and target contour characteristics of the image region through a multi-level attention calculation module to generate an enhanced feature tensor; A multi-dimensional parsing processing unit for multi-scale feature fusion is used to perform multi-scale convolution operations on the enhanced feature tensor, extract semantic and geometric features at different resolutions, and construct a three-dimensional feature space that contains global semantics and local details through cross-scale feature splicing and fusion operations; The obstacle recognition unit, based on dynamic threshold judgment, is used to use a preset obstacle feature template library in the three-dimensional feature space, adopt a sliding window search strategy, calculate feature similarity and spatial distance thresholds, identify obstacles in the environment, and obtain location coordinates, size information and confidence scores; The obstacle avoidance path planning and decision-making unit is used to generate multiple candidate obstacle avoidance paths in the flight-permitted area using a heuristic search algorithm based on the information output by the obstacle recognition unit, combined with the aircraft's current position, velocity vector, and a preset safety distance threshold; The motion control instruction generation unit is used to convert the optimal obstacle avoidance path selected by the obstacle avoidance path planning decision unit into the attitude adjustment parameters and speed control instructions of the aircraft, and calculate the motor speed adjustment amount and the servo deflection angle through the attitude solution algorithm and dynamic model.

2. The intelligent obstacle avoidance system for low-altitude aircraft according to claim 1, characterized in that: In the feature-enhanced attention unit with adaptive weight allocation, the improved attention mechanism adopts the following model formula: Among them, A ij represents the attention weight of the i-th position to the j-th position; f i 、f j 、f k are the feature vectors at the i, j, and k positions in the feature tensor respectively; W1 and W2 are learnable weight matrices; b is the bias vector; σ is the activation function; and N is the total number of positions in the feature tensor.

3. The intelligent obstacle avoidance system for low-altitude aircraft according to claim 1, characterized in that: In the multi-dimensional analysis processing unit of the multi-scale feature fusion, the feature fusion formula based on the binocular visual sensing multi-dimensional analysis model is: fusion =α1F s1 +α2F s2 +α3F s3 , where F fusion is the fused feature tensor; F s1 、F s2 、F s3 are the feature tensors extracted at different scales; α1, α2, and α3 are adaptive fusion weights, which are calculated using the following formula: Wherein, m = 1, 2, 3; l = 1, 2, 3; Represents the gradient norm of the nth feature map at the sth scale; M is the number of feature maps.

4. The intelligent obstacle avoidance system for low-altitude aircraft according to claim 1, characterized in that: In the obstacle recognition unit based on dynamic threshold determination, the improved obstacle recognition similarity calculation model formula is: Among them, S ij represents the similarity between the i-th area to be identified and the j-th obstacle feature template; f i p 、 are the eigenvalues ​​of the i-th area to be identified and the j-th obstacle feature template in the p-th feature dimension respectively; P is the total number of feature dimensions.

5. The intelligent obstacle avoidance system for low-altitude aircraft according to claim 1, characterized in that: In the obstacle avoidance path planning decision unit, the path cost evaluation function formula is: C path =β1D obstacle +β2L path +β3A change , where C path is the total path cost; D obstacle is the minimum distance between the path and the obstacle, which is calculated by the obstacle position information obtained by the binocular vision sensing multidimensional analytical model; L path is the path length; A change is the change in the direction of the path; β1, β2, and β3 are weight coefficients, which are dynamically adjusted after evaluating the importance of different factors through the improved attention mechanism.

6. The intelligent obstacle avoidance system for low-altitude aircraft according to claim 1, characterized in that: In the binocular visual perception unit based on spatiotemporal feature enhancement, the optical flow estimation improved model formula is: Among them, v x，y is the optical flow vector of the pixel point (x, y); I t , I t+1 are the images at time t and time t+1 respectively; m and n are the neighborhood pixel coordinate offsets; M and N are the neighborhood window sizes.

7. The intelligent obstacle avoidance system for low-altitude aircraft according to claim 1, characterized in that: The feature-enhanced attention unit with adaptive weight allocation also includes a secondary attention optimization module, and the model formula is: in, is the optimized attention weight; A ij is the attention weight obtained by the initial calculation; τ is the temperature parameter used to adjust the distribution of attention weight; k is the k-th accumulation, and N is the total number of accumulations.

8. The intelligent obstacle avoidance system for low-altitude aircraft according to claim 1, characterized in that: In the multi-dimensional parsing processing unit of the multi-scale feature fusion, a cross-channel attention mechanism is introduced, and its model formula is: in, is the feature tensor after channel optimization; F c is the feature tensor of the original c-th channel; F d is the feature tensor of the dth channel; D is the total number of channels; γ c is the channel retention coefficient; ω cd is the inter-channel attention weight, which is calculated by combining the improved attention mechanism with the correlation of different channel features in the binocular visual sensing multidimensional parsing model.

9. The intelligent obstacle avoidance system for low-altitude aircraft according to claim 1, characterized in that: In the obstacle avoidance path planning decision unit, a hierarchical search strategy is adopted, and the search range adjustment formula is: Among them, R search is the adjusted search range; R base As the basic search range; D nearest is the distance between the aircraft and the nearest obstacle, determined by the binocular vision sensing multi-dimensional analytical model; D safe is the preset safety distance; δ is the proportional coefficient.

10. The obstacle avoidance method of the intelligent obstacle avoidance system is characterized by: The following steps are involved: Step S1: using a binocular visual perception unit based on spatiotemporal feature enhancement to synchronously collect dual-channel image information of a low-altitude environment in an interlaced pixel sampling mode, and establishing pixel motion association between image frames through optical flow estimation to form multi-channel image tensor data; Step S2: Input the multi-channel image tensor data into the feature enhancement attention unit with adaptive weight allocation. Through the multi-level attention calculation module, the features of each channel and spatial position are dynamically weighted according to the texture complexity, edge gradient strength and target contour characteristics of the image area to generate an enhanced feature tensor; Step S3: The enhanced feature tensor is transmitted to a multi-dimensional parsing processing unit for multi-scale feature fusion, where a multi-scale convolution operation is performed to extract semantic and geometric features at different resolutions. A three-dimensional feature space containing global semantics and local details is constructed through cross-scale feature splicing and fusion operations. Step S4: The 3D feature space data is fed into an obstacle recognition unit based on dynamic threshold determination. Within the space, a preset obstacle feature template library is used, and a sliding window search strategy is adopted to identify obstacles in the environment by calculating feature similarity and spatial distance thresholds. The obstacles are then identified and their position coordinates, size information, and confidence scores are obtained. Step S5: The obstacle avoidance path planning decision unit generates multiple candidate obstacle avoidance paths in the flight allowed area using a heuristic search algorithm based on the information output by the obstacle recognition unit, combined with the aircraft's current position, velocity vector, and a preset safety distance threshold, and selects the optimal obstacle avoidance path using a path cost evaluation function. Step S6: The motion control instruction generation unit converts the optimal obstacle avoidance path into the attitude adjustment parameters and speed control instructions of the aircraft, calculates the motor speed adjustment amount and the servo deflection angle through the attitude solution algorithm and the dynamic model, and drives the aircraft to perform obstacle avoidance actions.

Images