An air-ground collaborative perception method and device based on reinforcement learning and a readable medium

Abstract

The application discloses an air-ground cooperative perception method and device based on reinforcement learning and a readable medium, and relates to the field of cooperative perception, which comprises the following steps: inputting first image data collected by a vehicle and second image data collected by an aircraft into a feature extraction module in a trained perception task classification and identification model, obtaining a first feature map and a second feature map of a current time slot, and transforming to obtain an aligned second feature map of the current time slot; constructing a state of the current time slot and inputting the state into a trained policy network to obtain a policy of the current time slot, and combining the state to segment the aligned second feature map to obtain a segmented second feature map of the current time slot, and performing fusion on the segmented second feature map of the current time slot and the first feature map of the current time slot through a feature fusion module to obtain a fusion feature map of the current time slot; and inputting the fusion feature map of the current time slot into a second feature decoding module to obtain a target class probability and a target spatial position of the current time slot. The application solves the problem of low perception accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of cooperative perception, in particular to an air-ground cooperative perception method based on reinforcement learning, an air-ground cooperative perception device and a readable medium. BACKGROUND

[0002] Current vehicle cooperative perception for automatic driving usually considers the interaction of laser radar point cloud, picture and motion state and other perception data between vehicles (vehicle-vehicle) and between vehicles and roadside infrastructure (vehicle-road), to solve the problems of vehicle visual blind area and limited perception distance. For example, based on the deep deterministic policy gradient algorithm, the cooperative vehicle is selected according to the available network bandwidth, the computing load of the vehicle, the communication distance and the historical cooperation record, etc. to reduce the communication overhead of data interaction. In addition, the deep Q learning algorithm is used to optimize the selection of transmission channel and data area according to the information such as the position of the surrounding vehicles, the channel gain and the network bandwidth, etc. to balance the communication bandwidth overhead and the perception performance gain.

[0003] However, in a complex traffic scene, due to the low deployment position of the laser radar and camera sensors of the vehicle and the roadside infrastructure, the perception field of view is easily blocked by large vehicles and surrounding buildings, resulting in failure to detect long-distance targets and obstacles in blocked areas, and unable to obtain comprehensive traffic environment information in time. In addition, under the conditions of bandwidth limitation and high dynamic topology change of vehicle communication, cooperative perception often faces problems such as data transmission delay and packet loss rate, which can cause target position deviation and missed detection and other perception errors. SUMMARY

[0004] The present application aims at the above-mentioned technical problems and proposes an air-ground cooperative perception method based on reinforcement learning, an air-ground cooperative perception device and a readable medium.

[0005] In a first aspect, the present application provides an air-ground cooperative perception method based on reinforcement learning, comprising the following steps:

[0006] A perception task classification and recognition model and a first feature decoding module are constructed and trained to obtain a trained perception task classification and recognition model and a trained first feature decoding module, and the perception task classification and recognition model comprises a feature extraction module, a feature fusion module and a second feature decoding module connected in sequence;

[0007] A strategy network and a value network are constructed and trained by using reinforcement learning to obtain a trained strategy network and a trained value network;

[0008] The first image data collected by the N cameras of the vehicle on the ground and the second image data collected by the N cameras of the aerial vehicle in the air are input into the trained perception task classification recognition model in the current time slot, and feature extraction is performed through the feature extraction module respectively to obtain the first feature map and the second feature map of the current time slot; the second feature map of the current time slot is mapped to the vehicle coordinate system through the pose transformation matrix to obtain the aligned second feature map of the current time slot; the state of the current time slot is constructed based on the aligned second feature map of the current time slot, the channel gain, the available bandwidth, and the transmission delay and the perception accuracy of the previous time slot, and the trained first feature decoding module is adopted;

[0009] The state of the current time slot is input into the trained strategy network to obtain the strategy of the current time slot, the aligned second feature map of the current time slot is segmented based on the strategy and the state of the current time slot to obtain the segmented second feature map of the current time slot, and the segmented second feature map of the current time slot is sent to the vehicle; the segmented second feature map of the current time slot and the first feature map of the current time slot are fused through the feature fusion module to obtain the fusion feature map of the current time slot, and the fusion feature map of the current time slot is input into the second feature decoding module to obtain the target class probability and the target spatial position of the current time slot.

[0010] As a preferred, the strategy network and the value network are constructed and trained by reinforcement learning to obtain the trained strategy network and the value network, which specifically includes:

[0011] S21, the transmission delay and the perception accuracy are initialized to obtain the transmission delay and the perception accuracy of the 0th time slot ; ;

[0012] S22, the strategy network and the value network are initialized to obtain the strategy network and the value network of the 1st time slot, and the empty experience replay buffer is set in the 1st time slot ;

[0013] S23, the first image data collected by the N cameras of the vehicle on the ground and the second image data collected by the N cameras of the aerial vehicle in the air are input into the trained perception task classification recognition model in the kth time slot, and feature extraction is performed through the feature extraction module to obtain the first feature map and the second feature map of the kth time slot, wherein represents the ith first feature map in the kth time slot, represents the ith second feature map in the kth time slot, H, W and C represent the length, width and channel number of the feature map respectively, and k is a positive integer; the second feature map of the kth time slot is mapped to the vehicle coordinate system through the pose transformation matrix Mapping to the vehicle coordinate system yields the aligned second feature map for the k-th time slot. As shown in the following formula:

[0014] ;

[0015] in, This represents the aligned second feature map of the i-th image in the k-th time slot;

[0016] Based on the aligned second feature map of the k-th time slot, channel gain, available bandwidth, and transmission delay and sensing accuracy of the (k-1)-th time slot, the state of the k-th time slot is constructed using the trained first feature decoding module. ;

[0017] S24, change the state of the kth time slot. The input is fed into the policy network of the k-th time slot to obtain the policy for the k-th time slot. Based on the strategy and state of the k-th time slot, the aligned second feature map of the k-th time slot is segmented to obtain the segmented second feature map of the k-th time slot. And sent to the vehicle, among which This represents the second feature map after segmentation in the k-th time slot; the second feature map after segmentation in the k-th time slot is fused with the first feature map in the k-th time slot through the feature fusion module to obtain the fused feature map of the k-th time slot; the fused feature map of the k-th time slot is input into the second feature decoding module to map the target category probability and target spatial location in the k-th time slot of the perception task.

[0018] S25, the perception accuracy of the k-th time slot is obtained based on the target category probability estimation of the k-th time slot. And estimate the transmission delay of the k-th time slot. The reward for the k-th time slot is calculated based on the sensing accuracy and transmission delay of the k-th time slot. ;

[0019] S26, Repeat step S23 to construct the state of the (k+1)th time slot and construct the empirical sample of the kth time slot. And write it to the experience playback buffer of the k-th time slot;

[0020] S27, Randomly sample empirical samples from the empirical playback buffer of the k-th time slot to obtain the empirical sample set. , the set of experience samples The state of the g-th time slot in the empirical sample of the g-th time slot Input the value network of the k-th time slot to obtain the corresponding value and the parameters of the value network of the k+1th time slot are obtained by updating the parameters of the value network of the kth time slot using a gradient descent algorithm to minimize the mean square error as shown in the following formula:

[0021] ;

[0022] wherein, is the cumulative return, denoted as , denotes the discount factor of the gth time slot, , and denote the policy of the gth time slot, the reward and the state of the g+1th time slot respectively, denotes the total number of experience samples in the experience sample set , denotes the parameter of the value network corresponding to the maximum value;

[0023] the parameters of the policy network of the k+1th time slot are obtained by updating the parameters of the policy network of the kth time slot using the following formula :

[0024] ;

[0025] wherein, is the advantage estimate, denoted as , is a truncation function, is a truncation coefficient, denotes the parameters of the policy network of the k+1th time slot, denotes the parameters of the policy network of the kth time slot, denotes the intermediate parameters of the policy network used in each update from the first update to the last update, denotes the input and outputs a probability distribution, denotes the input and outputs a probability distribution, denotes the intermediate parameters of the policy network corresponding to the maximum value;

[0026] S28, repeating steps S23-S27 until the policy network and the value network of the most recent time slot converge or the number of time slots reaches the maximum time slot number K.

[0027] As preferred, the state of the kth time slot is constructed based on the aligned second feature map of the kth time slot, channel gain, available bandwidth, and transmission delay and perception accuracy of the (k-1)th time slot, and by using the trained first feature decoding module , specifically comprising:

[0028] The aligned second feature map of the kth time slot is input into the trained first feature decoding module to obtain the confidence of each pixel coordinate in the aligned second feature map of the kth time slot, and a confidence matrix of the kth time slot is constructed;

[0029] The channel gain and available bandwidth between the aircraft and the vehicle in the kth time slot are estimated by using a preamble-based channel gain estimation method; the state of the kth time slot is constructed based on the confidence matrix of the kth time slot, channel gain, available bandwidth, transmission delay and perception accuracy of the previous time slot .

[0030] As preferred, the perception task includes target identification or target tracking; the feature extraction module includes four convolution modules connected in sequence, each convolution module including a first convolution layer, a batch normalization layer and a ReLU activation function layer connected in sequence; the feature fusion module includes a multi-head attention module, and the first feature decoding module and the second feature decoding module include four convolution networks connected in sequence, each convolution network splicing the feature at each position in the feature map output by the previous convolution network with the feature at the corresponding position in the feature map output by the feature extraction module to obtain a spliced feature map, and using a second convolution layer to halve the channel number of the spliced feature map to obtain the output feature map of each convolution network, and using two convolution branches on the output feature map of the last convolution network to classify the foreground and background categories respectively to output the target category probability and target spatial position.

[0031] As preferred, the aligned second feature map of the kth time slot is segmented based on the strategy and state of the kth time slot to obtain the segmented second feature map of the kth time slot, specifically comprising:

[0032] The strategy of the kth time slot includes a confidence selection threshold of the kth time slot, and a mask matrix corresponding to the i-th aligned second feature map in the kth time slot is constructed based on the confidence selection threshold of the kth time slot , as shown in the following formula:

[0033] ;

[0034] wherein, represents the selection state of the pixel coordinate (m, n) in the i-th aligned second feature map in the kth time slot, represents the confidence at the pixel coordinate (m, n) in the i-th aligned second feature map in the k-th time slot;

[0035] adopting the mask matrix corresponding to the i-th aligned second feature map in the k-th time slot adopting the i-th aligned second feature map in the k-th time slot performing segmentation to obtain the i-th segmented second feature map in the k-th time slot which only retains the region with the value of 1 .

[0036] As a preferred, the step S25 specifically comprises:

[0037] taking the maximum value in each row of the target category probability of the k-th time slot to construct a Q-dimensional vector, wherein, represents a set of real numbers, Q is the target number, C is the number of categories, and the average of all elements in the Q-dimensional vector is taken as the perception accuracy of the k-th time slot ; the time for sending the segmented second feature map of the k-th time slot from the aircraft to the vehicle is taken as the transmission delay of the k-th time slot ;

[0038] the reward of the k-th time slot is calculated by using the following formula :

[0039] ;

[0040] wherein, and are respectively the first weight coefficient and the second weight coefficient, is a penalty coefficient, represents an indicator function, the function value is 1 when the condition in the parentheses is true, otherwise it is 0; T is the maximum tolerable delay.

[0041] In a second aspect, the present application provides an air-ground collaborative perception device based on reinforcement learning, comprising:

[0042] a model construction module configured to construct a perception task classification identification model and a first feature decoding module and train to obtain a trained perception task classification identification model and a trained first feature decoding module, the perception task classification identification model comprising a feature extraction module, a feature fusion module and a second feature decoding module connected in turn;

[0043] a reinforcement learning module configured to construct a policy network and a value network and train by using reinforcement learning to obtain a trained policy network and a trained value network;

[0044] The state construction module is configured to acquire first image data collected by N cameras of a vehicle on the ground and second image data collected by N cameras of an aerial vehicle in a current time slot, input the first image data and the second image data into a trained perception task classification identification model, and perform feature extraction on the first image data and the second image data through a feature extraction module respectively to obtain a first feature map and a second feature map of the current time slot; map the second feature map of the current time slot to a vehicle coordinate system through a pose transformation matrix to obtain an aligned second feature map of the current time slot; and construct a state of the current time slot based on the aligned second feature map of the current time slot, channel gain, available bandwidth, transmission delay and perception accuracy of a previous time slot, and a trained first feature decoding module.

[0045] The perception module is configured to input the state of the current time slot into a trained strategy network to obtain a strategy of the current time slot, segment the aligned second feature map of the current time slot based on the strategy and the state of the current time slot to obtain a segmented second feature map of the current time slot, and send the segmented second feature map of the current time slot to the vehicle; fuse the segmented second feature map of the current time slot and the first feature map of the current time slot through a feature fusion module to obtain a fused feature map of the current time slot, and input the fused feature map of the current time slot into a second feature decoding module to obtain a target class probability and a target spatial position of the current time slot.

[0046] In a third aspect, the present application provides an electronic device, comprising one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any of the implementation manners of the first aspect.

[0047] In a fourth aspect, the present application provides a computer-readable storage medium, which stores a computer program, when the computer program is executed by a processor, the method described in any of the implementation manners of the first aspect is implemented.

[0048] In a fifth aspect, the present application provides a computer program product, comprising a computer program, when the computer program is executed by a processor, the method described in any of the implementation manners of the first aspect is implemented.

[0049] Compared with the prior art, the present application has the following beneficial effects:

[0050] (1) The air-ground collaborative perception method based on reinforcement learning introduced in the present application assists the ground vehicle in perceiving targets by means of an aerial vehicle, enhances the perception field of view by means of a higher sensor deployment position, improves the accuracy of target detection and other perception tasks, breaks through the space and perception field of view constraints based on the existing ground device collaboration, and further assists the vehicle in completing target detection and tracking and other perception tasks by means of the advantages of low-cost rapid deployment and aerial panoramic perception field of view of the unmanned aerial vehicle, thereby becoming an important technical supplement to make up for the shortcomings of car-car collaboration and car-road collaboration.

[0051] (2) The air-ground collaborative perception method based on reinforcement learning provided in the present application proposes a collaborative perception data interaction method based on reinforcement learning, optimizes the selection of the number of regions of perception data based on available bandwidth, channel gain between the aircraft and the vehicle, the amount of perception data, the confidence of the perception data, and historical transmission delay, and improves the perception speed under the condition of limited bandwidth and fast time-varying channel. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0053] Figure 1 The flowchart of the air-ground collaborative perception method based on reinforcement learning of the embodiments of the present application is shown.

[0054] Figure 2 The schematic diagram of the air-ground collaborative perception device based on reinforcement learning of the embodiments of the present application is shown.

[0055] Figure 3 The hardware structure schematic diagram of the electronic device provided in the embodiments of the present application is shown. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solutions and advantages of the present application more clear, the present application will be further described in detail below with reference to the drawings. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0057] Figure 1 A kind of air-ground collaborative perception method based on reinforcement learning provided in the embodiments of the present application is shown, including the following steps:

[0058] S1, a perception task classification identification model and a first feature decoding module are constructed and trained, to obtain a trained perception task classification identification model and a trained first feature decoding module, the perception task classification identification model includes a feature extraction module, a feature fusion module and a second feature decoding module connected in turn.

[0059] In a specific embodiment, the perception task includes target recognition or target tracking; the feature extraction module includes four convolutional modules connected in sequence, each convolutional module including a first convolutional layer, a batch normalization layer and a ReLU activation function layer connected in sequence; the second feature decoding module and the first feature decoding module include a multi-head attention module, the feature decoding module includes four convolutional networks connected in sequence, each convolutional network concatenates the features at each position in the feature map output by the previous convolutional network with the features at the corresponding positions in the feature map output by the feature extraction module to obtain a concatenated feature map, and uses the second convolutional layer to halve the number of channels in the concatenated feature map to obtain the output feature map of each convolutional network, and uses two convolutional branches to classify the foreground and background categories respectively, outputting the target category probability and the target spatial location.

[0060] Specifically, in the embodiments of this application, the perception task classification and recognition model is first trained to obtain a trained perception task classification and recognition model. The feature extraction module, feature fusion module, and second feature decoding module included in the trained perception task classification and recognition model participate in the calculation during the subsequent reinforcement learning process. In one embodiment of this application, the feature extraction module includes four convolutional modules consisting of a two-dimensional first convolutional layer, a batch normalization layer, and a ReLU activation function layer. The feature fusion module adopts a multi-head attention module. The calculation process of the multi-head attention module is existing technology and will not be described in detail here. The second feature decoding module is used for feature mapping. In addition, the embodiments of this application also use a trained first feature decoding module. The second feature decoding module has the same structure as the first feature decoding module. In other embodiments, the feature extraction module, feature fusion module, second feature decoding module, and first feature decoding module can adopt other structures and are not limited to a specific type of network structure, as long as they can complete the necessary functions of the perception task.

[0061] S2. Construct a policy network and a value network and train them using reinforcement learning to obtain the trained policy network and value network.

[0062] In a specific embodiment, a policy network and a value network are constructed and trained using reinforcement learning to obtain trained policy and value networks, specifically including:

[0063] S21, initialize the transmission delay and sensing accuracy to obtain the transmission delay of the 0th time slot. and perception accuracy ;

[0064] S22, initialize the policy network and value network to obtain the policy network and value network for the first time slot, and empty the experience replay buffer in the first time slot. ;

[0065] S23, the first image data collected by the N cameras of the vehicle on the ground and the second image data collected by the N cameras of the aerial vehicle in the kth time slot are input into the trained perception task classification recognition model, and feature extraction is performed through the feature extraction module respectively to obtain the first feature map and the second feature map in the kth time slot , wherein represents the i-th first feature map in the kth time slot, H, W and C represent the length, width and channel number of the feature map respectively, and k is a positive integer; The second feature map in the kth time slot is mapped to the vehicle coordinate system through the pose transformation matrix , as shown in the following formula:

[0066] ;

[0067] , wherein represents the i-th aligned second feature map in the kth time slot;

[0068] The state of the kth time slot is constructed based on the aligned second feature map of the kth time slot, the channel gain, the available bandwidth, the transmission delay and the perception accuracy of the k-1th time slot, and the trained first feature decoding module ;

[0069] S24, the state of the kth time slot is input into the policy network of the kth time slot to obtain the policy of the kth time slot , the aligned second feature map of the kth time slot is segmented based on the policy and the state of the kth time slot to obtain the segmented second feature map of the kth time slot , and is sent to the vehicle, wherein represents the i-th segmented second feature map in the kth time slot; the segmented second feature map of the kth time slot and the first feature map of the kth time slot are fused through the feature fusion module to obtain the fusion feature map of the kth time slot, and the fusion feature map of the kth time slot is input into the second feature decoding module to obtain the target class probability and target spatial position in the kth time slot of the perception task through mapping;

[0070] S25, the perception accuracy of the kth time slot is estimated based on the target class probability of the kth time slot , and the transmission delay of the kth time slot is estimated ; the reward of the kth time slot is calculated based on the perception accuracy and the transmission delay of the kth time slot ; ​​

[0071] S26, repeat step S23 to build the state of the k+1th time slot, build the experience sample of the kth time slot , and write into the experience replay buffer of the kth time slot;

[0072] S27, randomly sample experience samples from the experience replay buffer of the kth time slot to obtain an experience sample set , and input the state of the gth time slot in the experience sample of the gth time slot in the experience sample set into the value network of the kth time slot to obtain the corresponding value , and update the parameters of the value network of the kth time slot using the gradient descent algorithm to minimize the mean square error to obtain the parameters of the value network of the k+1th time slot , as shown in the following formula:

[0073] ;

[0074] wherein, is the cumulative return, expressed as , is the discount coefficient of the gth time slot, , and respectively represent the policy of the gth time slot, the reward and the state of the g+1th time slot, represents the total number of experience samples in the experience sample set , and represents the parameters of the value network corresponding to the maximum value;

[0075] update the parameters of the policy network of the kth time slot using the following formula to obtain the parameters of the policy network of the k+1th time slot :

[0076] ;

[0077] wherein, is the advantage estimate, expressed as , is a truncation function, is a truncation coefficient, represents the parameters of the policy network of the k+1th time slot, represents the parameters of the policy network of the kth time slot, represents the intermediate parameters of the policy network used in each update from the first update to the last update, represents the probability distribution of inputting and outputting when using the intermediate parameters of the policy network, ​denotes the input when using the parameter of the policy network of the kth time slot and outputs a probability distribution of

[0078] S28, repeating steps S23-S27 until the policy network of the most recent time slot converges with the value network or the number of time slots reaches the maximum number of time slots K.

[0079] Specifically, the process of reinforcement learning in the embodiments of the present application is learned through time slot iteration. The architecture of the policy network and the value network used in the reinforcement learning process of the embodiments of the present application should be related to the dimension of the state and the dimension of the policy, and the specific network form can not be limited to a fully connected neural network, a convolutional neural network, etc., but the neuron dimension of the input and output should be related to the dimension of the current state and the dimension of the policy.

[0080] In one embodiment, the policy network and the value network are taken as an example of a fully connected neural network, which can be a 3-layer multi-layer perceptron, wherein the number of neurons in the first layer is Nxm*n+4, the number of output neurons is N, and the number of intermediate neurons can be selected (such as 128, 256, etc.).

[0081] In a specific embodiment, based on the aligned second feature map of the kth time slot, the channel gain, the available bandwidth, and the transmission delay and the perception accuracy of the (k-1)th time slot, and using the trained first feature decoding module, the state of the kth time slot is constructed , specifically comprising:

[0082] The aligned second feature map of the kth time slot is input into the trained first feature decoding module to obtain the confidence of each pixel coordinate in the aligned second feature map of the kth time slot, and a confidence matrix of the kth time slot is constructed.

[0083] The channel gain and the available bandwidth between the aircraft and the vehicle in the kth time slot are estimated using a preamble-based channel gain estimation method; based on the confidence matrix of the kth time slot, the channel gain, the available bandwidth, the transmission delay and the perception accuracy of the previous time slot, the state of the kth time slot is constructed .

[0084] In a specific embodiment, the aligned second feature map of the kth time slot is segmented based on the policy and the state of the kth time slot to obtain the segmented second feature map of the kth time slot, specifically comprising:

[0085] The policy of the kth time slot includes a confidence selection threshold of the kth time slot, and based on the confidence selection threshold of the kth time slot, a mask matrix corresponding to the i-th aligned second feature map in the kth time slot is constructed , as shown in the following formula:

[0086] ;

[0087] wherein, represents the selection state of the pixel coordinate (m, n) in the aligned second feature map of the ith image in the kth time slot, represents the confidence of the pixel coordinate (m, n) in the aligned second feature map of the ith image in the kth time slot.

[0088] The mask matrix corresponding to the aligned second feature map of the ith image in the kth time slot is adopted The aligned second feature map of the ith image in the kth time slot is segmented to obtain the segmented second feature map of the ith image in the kth time slot which only retains the region with value of 1. .

[0089] In specific embodiments, step S25 specifically comprises:

[0090] The maximum value in each row of the target class probability of the kth time slot is taken out to construct a Q-dimensional vector, wherein, represents a set of real numbers, Q is the target number, C is the number of categories, and the average value of all elements in the Q-dimensional vector is taken as the perception accuracy of the kth time slot ; the time for sending the segmented second feature map of the kth time slot from the aircraft to the vehicle is taken as the transmission delay of the kth time slot ;

[0091] The reward of the kth time slot is calculated by using the following formula :

[0092] ;

[0093] wherein, and are the first weight coefficient and the second weight coefficient respectively, is a penalty coefficient, represents an indicator function, the function value is 1 when the condition in the parentheses is true, otherwise it is 0; T is the maximum tolerable delay.

[0094] Embodiments of the present application take a drone as an example for illustration, and the parameters of the policy network and the value network are adjusted and updated through the reinforcement learning process, and the process of reinforcement learning is as follows:

[0095] 1. Initialize system parameters, including the initialized transmission delay and the perception accuracy .

[0096] ​​2. The construction parameters are the policy network , the value network with parameters , randomly initialize the parameters of the two networks and , and empty the experience replay buffer .

[0097] 3. In the kth (k≤K) time slot, the vehicle and the UAV carrying N high-definition cameras collect first image data and second image data , respectively, where represents the first image taken by the vehicle of the ith high-definition camera in the kth time slot, i.e., the ith first image, represents the second image taken by the UAV of the ith high-definition camera in the kth time slot, i.e., the ith second image, and further using the feature extraction module in the trained perception task classification recognition model to extract the first feature map and the second feature map corresponding to the two-dimensional bird's eye view (BEV) of the first image data and the second image data in the kth time slot, respectively, where , , H, W, and C represent the length, width, and channel number of the feature map, respectively.

[0098] 4. Map the second feature map of the kth time slot to the vehicle coordinate system one-to-one based on the pose transformation matrix to obtain the aligned second feature map of the kth time slot, 1≤i≤N.

[0099] 5. Input the aligned second feature map of the kth time slot to the trained first feature decoding module to obtain the confidence matrix of the ith aligned second feature map in the kth time slot, where the element in the confidence matrix is the confidence obtained by inputting the feature of each pixel coordinate in the ith aligned second feature map in the kth time slot to the trained first feature decoding module. The confidence matrix in the kth time slot is a matrix obtained by concatenating the confidence matrices of the N aligned second feature maps in the kth time slot, denoted as .

[0100] 6. Obtain the channel gain between the UAV and the vehicle in the kth time slot using the preamble-based channel estimation method, and obtain the available bandwidth in the kth time slot., which is a classical channel gain estimation method, the details of which are not described here. The available bandwidth is usually read, combined with the confidence matrix of the kth time slot , the transmission delay of the k-1th time slot and the perception accuracy to build the state of the kth time slot, as shown in the following formula:

[0101] .

[0102] 7. The state of the kth time slot is input into the policy network to obtain the policy of the kth time slot, i.e. the confidence selection threshold , which is used to dynamically select the data area in the aligned second feature map of the kth time slot with a confidence greater than . Specifically, first, the mask matrix of the ith second feature map in the kth time slot is constructed, where the element represents the selection state of the pixel coordinate (m, n) in the confidence matrix of the ith second feature map in the kth time slot; the ith second feature map in the kth time slot is segmented by the mask matrix , only the feature map area with a value of 1 in the mask matrix is retained, to obtain the segmented feature map of the kth time slot containing only the high-confidence area, and the segmented second feature map of the kth time slot is sent to the vehicle.

[0103] 8. The first feature map of the kth time slot is fused with the segmented second feature map to generate the fusion feature map of the kth time slot, and the feature decoding module in the trained perception task classification recognition model is further used to map the fusion feature map of the kth time slot to obtain the target class probability and the target spatial position , where Q is the target number and C is the number of classes.

[0104] 9. The vehicle estimates the perception accuracy of the kth time slot based on the target class probability of the kth time slot, evaluates the transmission delay of the segmented second feature map of the kth time slot, and feeds back the perception accuracy and transmission delay of the kth time slot to the UAV. The perception accuracy is estimated by the target class probability, and the target class probability The largest number in each row is used to construct a Q-dimensional vector, and the average of these Q numbers is taken as the perception accuracy. The transmission latency is evaluated based on timestamps. When the transmitter (UAV) sends the segmented second feature map of the k-th time slot, it adds a first timestamp. Then, the receiver (vehicle) subtracts the first timestamp from the second timestamp when it receives the segmented second feature map of the k-th time slot to obtain the transmission latency of the k-th time slot.

[0105] 10. Perception accuracy of UAV based on the k-th time slot and transmission delay The reward for the k-th time slot is calculated using a reward function.

[0106] 11. The drone moves to the next state. Constructing experience samples And write to the experience playback buffer .

[0107] 12. From the experience replay buffer Medium-sampling empirical sample set , the set of experience samples The g-th state Input Value Network Obtain the value of the kth time slot The gradient descent algorithm is used to minimize the mean square error and update the parameters of the value network.

[0108] 13. Update the policy network The parameters.

[0109] 14. Repeat steps 3-13 until the policy network and value network converge or the number of time slots reaches the maximum number of time slots K. The above process, based on reinforcement learning, operates on a time slot-by-time basis, continuously updating the policy network and value network during the learning process. Convergence is ultimately determined by whether the change in the policy network's parameters is less than a threshold. ,in This represents the threshold.

[0110] S3: In the current time slot, the first image data collected by N cameras on the ground vehicle and the second image data collected by N cameras on the air vehicle are acquired and input into the trained perception task classification and recognition model. The data are then processed by the feature extraction module to extract features, resulting in the first feature map and the second feature map of the current time slot. The second feature map of the current time slot is mapped to the vehicle coordinate system through a pose transformation matrix to obtain the aligned second feature map of the current time slot. Based on the aligned second feature map of the current time slot, the channel gain, the available bandwidth, and the transmission delay and perception accuracy of the previous time slot, the state of the current time slot is constructed using the trained first feature decoding module.

[0111] Specifically, in the embodiments of the present application, the trained perception task classification recognition model and the trained strategy network are applied to the perception task classification process of the first image data and the second image data collected in any subsequent time slot, that is, the first image data and the second image data obtained in the current time slot are input into the trained perception task classification recognition model, and the first feature map and the second feature map of the current time slot are extracted through the feature extraction module respectively, then the second feature map of the current time slot is mapped to the vehicle coordinate system through the pose transformation matrix to obtain the aligned second feature map of the current time slot; the state of the current time slot is constructed based on the aligned second feature map of the current time slot, the channel gain, the available bandwidth, the transmission delay and the perception accuracy of the previous time slot, and the trained first feature decoding module, which is the same as the process of constructing the state of the kth time slot based on the aligned second feature map of the kth time slot, the channel gain, the available bandwidth, the transmission delay and the perception accuracy of the (k-1) th time slot, and the trained first feature decoding module in the reinforcement learning process, and will not be described here.

[0112] S4, inputting the state of the current time slot into the trained strategy network to obtain the strategy of the current time slot, segmenting the aligned second feature map of the current time slot based on the strategy and the state of the current time slot to obtain the segmented second feature map of the current time slot, and sending the segmented second feature map of the current time slot to the vehicle; fusing the segmented second feature map of the current time slot and the first feature map of the current time slot through the feature fusion module to obtain the fusion feature map of the current time slot, and inputting the fusion feature map of the current time slot into the second feature decoding module to map to obtain the target class probability and the target spatial position of the current time slot.

[0113] Specifically, after the state of the current time slot is constructed, it is input into the trained strategy network to obtain the strategy of the current time slot, that is, the confidence selection threshold of the current time slot, based on which the mask matrix of the current time slot can be constructed, and the aligned second feature map of the current time slot can be segmented based on the mask matrix to obtain the segmented second feature map of the current time slot. Wherein, the process of inputting the state of the current time slot into the trained strategy network to obtain the strategy of the current time slot, and segmenting the aligned second feature map of the current time slot based on the strategy and the state of the current time slot to obtain the segmented second feature map of the current time slot, is the same as the process of segmenting the aligned second feature map of the kth time slot based on the strategy and the state of the kth time slot to obtain the segmented second feature map of the kth time slot, and will not be described here.

[0114] Further, the segmented second feature map of the current time slot is fused with the first feature map of the current time slot through a feature fusion module to obtain a fusion feature map of the current time slot, and the fusion feature map of the current time slot is input into a second feature decoding module to obtain a target category probability and a target spatial position of the current time slot through decoding.

[0115] Further reference Figure 2 , as an implementation of the method shown in the above figures, the present application provides an embodiment of an air-ground collaborative perception device based on reinforcement learning, which corresponds to the method embodiment shown in Figure 1 , and the device can be applied to various electronic devices.

[0116] The embodiment of the present application provides an air-ground collaborative perception device based on reinforcement learning, which comprises:

[0117] A model construction module 1 is configured to construct a perception task classification and recognition model and a first feature decoding module and train to obtain a trained perception task classification and recognition model and a trained first feature decoding module, wherein the perception task classification and recognition model comprises a feature extraction module, a feature fusion module and a second feature decoding module connected in sequence;

[0118] A reinforcement learning module 2 is configured to construct a policy network and a value network and train by using reinforcement learning to obtain a trained policy network and a trained value network;

[0119] A state construction module 3 is configured to input first image data collected by N cameras of a vehicle on the ground and second image data collected by N cameras of an aircraft in the air to the trained perception task classification and recognition model and perform feature extraction through the feature extraction module respectively to obtain a first feature map and a second feature map of a current time slot; map the second feature map of the current time slot to a vehicle coordinate system through a pose transformation matrix to obtain an aligned second feature map of the current time slot; and construct a state of the current time slot based on the aligned second feature map of the current time slot, channel gain, available bandwidth, transmission delay and perception accuracy of a previous time slot and the trained first feature decoding module.

[0120] A perception module 4 is configured to input the state of the current time slot into the trained policy network to obtain a policy of the current time slot, segment the aligned second feature map of the current time slot based on the policy and the state of the current time slot to obtain a segmented second feature map of the current time slot, and send the segmented second feature map of the current time slot to the vehicle; fuse the segmented second feature map of the current time slot with the first feature map of the current time slot through a feature fusion module to obtain a fusion feature map of the current time slot, input the fusion feature map of the current time slot into a second feature decoding module, and map to obtain a target category probability and a target spatial position of the current time slot.

[0121] Figure 3 A hardware structure schematic diagram of an electronic device provided by the embodiment of the present application is shown in the figure. As shown in the figure, the electronic device of the embodiment comprises a processor 301 and a memory 302; wherein the memory 302 is used for storing computer execution instructions; the processor 301 is used for executing the computer execution instructions stored by the memory to realize each step executed by the electronic device in the above embodiment. For details, please refer to the related description in the foregoing method embodiment. Figure 3

[0122] Optionally, the memory 302 can be independent or integrated with the processor 301.

[0123] When the memory 302 is independently arranged, the electronic device further comprises a bus 303 for connecting the memory 302 and the processor 301.

[0124] The embodiment of the present application further provides a computer storage medium, and the computer storage medium stores computer execution instructions. When the processor 301 executes the computer execution instructions, the method as above is realized.

[0125] The embodiment of the present application further provides a computer program product, and the computer program product comprises a computer program. When the computer program is executed by the processor 301, the method as above is realized.

[0126] In the embodiments provided by the present application, it should be understood that the disclosed device and method can be implemented in other ways. For example, the device embodiments described above are only schematic. The division of the modules is only a logical function division. There can be another division manner in actual implementation, for example, a plurality of modules can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the modules shown or discussed can be indirect coupling or communication connection through some interfaces, devices or modules, and can be electrical, mechanical or in other forms.

[0127] The modules illustrated as separate components can or can not be physically separate, and the components illustrated as modules can or can not be physical units, i.e. can be located in one place, or can be distributed on a plurality of network units. Some or all of the modules can be selected according to actual needs to implement the embodiment scheme.

[0128] In addition, each functional module in each embodiment of the present application can be integrated in one processing unit, or each module can be physically present independently, or two or more modules can be integrated in one unit. The unit formed by the above modules can be realized in the form of hardware, or in the form of hardware plus software function unit.​

[0129] The integrated modules implemented in the form of software function modules can be stored in a computer readable storage medium. The software function modules are stored in a storage medium and include a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or the processor 301 to execute part of the steps of the method of various embodiments of the present application.

[0130] It should be understood that the processor 301 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), etc. The general-purpose processor can be a microprocessor or the processor 301 can also be any conventional processor 301, etc. The steps of the method disclosed in combination with the application can be directly embodied as the execution of the processor 301 in hardware, or executed by a combination of hardware and software modules in the processor 301.

[0131] The memory 302 can include a high-speed RAM memory, and can also include a non-volatile storage NVM, for example at least one disk memory, and can also be a U disk, a mobile hard disk, a read-only memory, a magnetic disk or an optical disk, etc.

[0132] The bus 303 can be an industry standard architecture (ISA), a peripheral component interconnect (PCI) bus, or an extended industry standard architecture (EISA) bus, etc. The bus 303 can be divided into an address bus, a data bus, a control bus, etc. For the sake of convenience, the bus 303 in the drawings of the present application does not limit to only one bus 303 or one type of bus 303.

[0133] The storage medium can be realized by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0134] An exemplary storage medium is coupled to the processor 301 such that the processor 301 can read information from, and can write information to, the storage medium. Of course, the storage medium can be a part of the processor 301. Consistent with the teachings provided herein, the processor 301 and the storage medium can be located in a ASIC. The teachings provided herein can be used in other electronic devices such as a computer system, a consumer electronics device, or any other device that includes a processor and a storage medium.

[0135] Those skilled in the art can understand that all or part of the steps of the foregoing method embodiments can be completed by a program instruction related to hardware. The foregoing program can be stored in a computer readable storage medium. When the program is executed, the steps of the foregoing method embodiments are executed; and the foregoing storage medium includes: ROM, RAM, disk or disc, and various media that can store program codes.

[0136] Finally, it should be noted that the above embodiments are merely used to illustrate, but not to limit the technical solutions of the present application; even though the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A reinforcement learning-based air-ground cooperative perception method, characterized in that, Includes the following steps: A perception task classification and recognition model and a first feature decoding module are constructed and trained to obtain a trained perception task classification and recognition model and a trained first feature decoding module. The perception task classification and recognition model includes a feature extraction module, a feature fusion module and a second feature decoding module connected in sequence. Constructing a policy network and a value network and training them using reinforcement learning yields trained policy and value networks. Specifically, this involves acquiring first image data from N cameras on a ground-based vehicle and second image data from N cameras on an airborne aircraft in the k-th time slot. These images are then input into the trained perception task classification and recognition model, and feature extraction is performed by the feature extraction module to obtain the first feature map for the k-th time slot. Second feature map ,in This represents the i-th first feature map in the k-th time slot. This represents the i-th second feature map in the k-th time slot, where H, W, and C represent the length, width, and number of channels of the feature map, respectively, and k is a positive integer. The second feature map of the k-th time slot is transformed by the pose transformation matrix. Mapping to the vehicle coordinate system yields the aligned second feature map for the k-th time slot. As shown in the following formula: ； in, This represents the aligned second feature map of the i-th image in the k-th time slot; Based on the aligned second feature map of the k-th time slot, channel gain, available bandwidth, and transmission delay and sensing accuracy of the (k-1)-th time slot, the state of the k-th time slot is constructed using the trained first feature decoding module. ; The state of the kth time slot The input is fed into the policy network of the k-th time slot to obtain the policy for the k-th time slot. Based on the strategy and state of the k-th time slot, the aligned second feature map of the k-th time slot is segmented to obtain the segmented second feature map of the k-th time slot. and sent to the vehicle, wherein The second feature map after segmentation in the k-th time slot is represented by the i-th segmented feature map. The second feature map after segmentation in the k-th time slot is fused with the first feature map in the k-th time slot through the feature fusion module to obtain the fused feature map of the k-th time slot. The fused feature map of the k-th time slot is input into the second feature decoding module to map the target category probability and target spatial location in the k-th time slot of the perception task. The perception accuracy of the k-th time slot is obtained based on the target category probability estimation of the k-th time slot. And estimate the transmission delay of the k-th time slot. The reward for the k-th time slot is calculated based on the sensing accuracy and transmission delay of the k-th time slot. ; In the current time slot, the first image data acquired by N cameras on a ground vehicle and the second image data acquired by N cameras on an airborne aircraft are input into the trained perception task classification and recognition model. These images are then processed by the feature extraction module to extract features, resulting in a first feature map and a second feature map for the current time slot. The second feature map of the current time slot is mapped to the vehicle coordinate system using a pose transformation matrix to obtain an aligned second feature map for the current time slot. Based on the aligned second feature map, channel gain, available bandwidth, transmission delay, and perception accuracy of the previous time slot, and using the trained first feature decoding module, the state of the current time slot is constructed. The state of the current time slot is input into the trained policy network to obtain the policy for the current time slot. Based on the policy and state of the current time slot, the aligned second feature map of the current time slot is segmented to obtain the segmented second feature map of the current time slot, which is then sent to the vehicle. The segmented second feature map of the current time slot is fused with the first feature map of the current time slot through the feature fusion module to obtain the fused feature map of the current time slot. The fused feature map of the current time slot is input into the second feature decoding module to map the target category probability and target spatial location of the current time slot.

2. The air-ground cooperative sensing method based on reinforcement learning according to claim 1, characterized in that, Construct a policy network and a value network and train them using reinforcement learning to obtain the trained policy network and value network, specifically including: S21, initialize the transmission delay and sensing accuracy to obtain the transmission delay of the 0th time slot. and perception accuracy ; S22, initialize the policy network and value network to obtain the policy network and value network for the first time slot, and empty the experience replay buffer in the first time slot. ; S23, in the k-th time slot, the first image data collected by N cameras on the ground vehicles and the second image data collected by N cameras on the air vehicles are acquired and input into the trained perception task classification and recognition model, and feature extraction is performed by the feature extraction module to obtain the first feature map of the k-th time slot. Second feature map ,in This represents the i-th first feature map in the k-th time slot. This represents the i-th second feature map in the k-th time slot, where H, W, and C represent the length, width, and number of channels of the feature map, respectively, and k is a positive integer. The second feature map of the k-th time slot is transformed by the pose transformation matrix. Mapping to the vehicle coordinate system yields the aligned second feature map for the k-th time slot. As shown in the following formula: ； in, This represents the aligned second feature map of the i-th image in the k-th time slot; Based on the aligned second feature map of the k-th time slot, channel gain, available bandwidth, and transmission delay and sensing accuracy of the (k-1)-th time slot, the state of the k-th time slot is constructed using the trained first feature decoding module. ; S24, change the state of the kth time slot. The input is fed into the policy network of the k-th time slot to obtain the policy for the k-th time slot. Based on the strategy and state of the k-th time slot, the aligned second feature map of the k-th time slot is segmented to obtain the segmented second feature map of the k-th time slot. and sent to the vehicle, wherein The second feature map after segmentation in the k-th time slot is represented by the i-th segmented feature map. The second feature map after segmentation in the k-th time slot is fused with the first feature map in the k-th time slot through the feature fusion module to obtain the fused feature map of the k-th time slot. The fused feature map of the k-th time slot is input into the second feature decoding module to map the target category probability and target spatial location in the k-th time slot of the perception task. S25, the perception accuracy of the k-th time slot is obtained based on the target category probability estimation of the k-th time slot. And estimate the transmission delay of the k-th time slot. The reward for the k-th time slot is calculated based on the sensing accuracy and transmission delay of the k-th time slot. ; S26, Repeat step S23 to construct the state of the (k+1)th time slot and construct the empirical sample of the kth time slot. And write it to the experience playback buffer of the k-th time slot; S27, Randomly sample empirical samples from the empirical playback buffer of the k-th time slot to obtain the empirical sample set. The set of experience samples The state of the g-th time slot in the empirical sample of the g-th time slot Input the value network of the k-th time slot to obtain the corresponding value The gradient descent algorithm is used to minimize the mean square error and update the parameters of the value network in the k-th time slot, thus obtaining the parameters of the value network in the (k+1)-th time slot. As shown in the following formula: ； in, For cumulative returns, it is represented as , This represents the discount factor for the g-th time slot. , and Let these represent the policy, reward, and state of the g-th time slot, respectively. Represents the set of empirical samples The total number of empirical samples in the data. This represents the parameters of the value network corresponding to the maximum value. The parameters of the policy network in the k-th time slot are updated using the following formula to obtain the parameters of the policy network in the (k+1)-th time slot. : ； in, For the advantage estimate, it is expressed as , This is a truncation function. The cutoff factor is , This represents the parameters of the policy network in the (k+1)th time slot. This represents the parameters of the policy network in the k-th time slot. This represents the intermediate parameters used in each update of the policy network from the first update to the last update. This indicates the input when using intermediate parameters of the policy network. and output The probability distribution, This indicates the input parameters when using the policy network in the k-th time slot. and output The probability distribution, This represents the intermediate parameters of the policy network corresponding to the maximum value. S28. Repeat steps S23-S27 until the policy network and value network of the most recent time slot converge or the number of time slots reaches the maximum number of time slots K.

3. The air-ground cooperative sensing method based on reinforcement learning according to claim 2, characterized in that, Based on the aligned second feature map of the k-th time slot, channel gain, available bandwidth, and transmission delay and sensing accuracy of the (k-1)-th time slot, the state of the k-th time slot is constructed using the trained first feature decoding module. Specifically, it includes: The aligned second feature map of the k-th time slot is input into the trained first feature decoding module to obtain the confidence score of each pixel coordinate in the aligned second feature map of the k-th time slot, and to construct the confidence score matrix of the k-th time slot. The channel gain and available bandwidth between the aircraft and the vehicle in the k-th time slot are estimated using a preamble-based channel gain estimation method; the confidence matrix of the k-th time slot is then used. Channel gain Available bandwidth Transmission delay of the (k-1)th time slot and perception accuracy Construct the state of the k-th time slot .

4. The air-ground cooperative sensing method based on reinforcement learning according to claim 3, characterized in that, The perception task includes target recognition or target tracking; the feature extraction module includes four convolutional modules connected in sequence, each convolutional module including a first convolutional layer, a batch normalization layer, and a ReLU activation function layer connected in sequence; the feature fusion module includes a multi-head attention module, the first feature decoding module and the second feature decoding module include four convolutional networks connected in sequence, each convolutional network concatenates the features at each position in the feature map output by the previous convolutional network with the features at the corresponding positions in the feature map output by the feature extraction module to obtain a concatenated feature map, and uses a second convolutional layer to halve the number of channels in the concatenated feature map to obtain the output feature map of each convolutional network, and uses two convolutional branches to classify the foreground and background categories respectively, outputting the target category probability and the target spatial location.

5. The air-ground cooperative sensing method based on reinforcement learning according to claim 3, characterized in that, Based on the policy and state of the k-th time slot, the aligned second feature map of the k-th time slot is segmented to obtain the segmented second feature map of the k-th time slot, specifically including: The strategy for the k-th time slot includes a confidence selection threshold for the k-th time slot, and a mask matrix corresponding to the aligned second feature map in the k-th time slot is constructed based on the confidence selection threshold for the k-th time slot. As shown in the following formula: ； in, This represents the selection state at pixel coordinates (m, n) in the i-th aligned second feature map within the k-th time slot. This represents the confidence level at pixel coordinates (m, n) in the i-th aligned second feature map within the k-th time slot; The mask matrix corresponding to the i-th aligned second feature map in the k-th time slot is used. The aligned second feature map in the k-th time slot Segmentation is performed to obtain only the retained The second feature map after segmentation in the k-th time slot of the region with a value of 1 .

6. The air-ground cooperative sensing method based on reinforcement learning according to claim 2, characterized in that, Step S25 specifically includes: Extract the target category probability of the k-th time slot Construct a Q-dimensional vector from the largest number in each row, where... Let Q represent the set of real numbers, C be the number of targets, and C be the number of categories. The mean of all elements in the Q-dimensional vector is taken as the sensing accuracy for the k-th time slot. The time taken for the second feature map after the k-th time slot is segmented and sent from the aircraft to the vehicle is taken as the transmission delay of the k-th time slot. ; The reward for the k-th time slot is calculated using the following formula. : ； in, and These are the first weighting coefficient and the second weighting coefficient, respectively. The penalty coefficient is... This indicates an indicator function; the function value is 1 when the condition in parentheses is true, and 0 otherwise; T is the maximum tolerable delay.

7. A reinforcement learning-based air-to-ground cooperative sensing device, employing the reinforcement learning-based air-to-ground cooperative sensing method according to any one of claims 1-6, characterized in that, include: The model building module is configured to build and train a perception task classification and recognition model and a first feature decoding module to obtain a trained perception task classification and recognition model and a trained first feature decoding module. The perception task classification and recognition model includes a feature extraction module, a feature fusion module and a second feature decoding module connected in sequence. The reinforcement learning module is configured to build a policy network and a value network and train them using reinforcement learning to obtain trained policy networks and value networks. The state construction module is configured to acquire first image data collected by N cameras on a ground vehicle and second image data collected by N cameras on an airborne aircraft in the current time slot, input them into the trained perception task classification and recognition model, and extract features through the feature extraction module to obtain a first feature map and a second feature map for the current time slot; map the second feature map of the current time slot to the vehicle coordinate system through a pose transformation matrix to obtain an aligned second feature map of the current time slot; and construct the state of the current time slot based on the aligned second feature map of the current time slot, channel gain, available bandwidth, transmission delay and perception accuracy of the previous time slot, and using the trained first feature decoding module. The perception module is configured to input the state of the current time slot into the trained policy network to obtain the policy for the current time slot; segment the aligned second feature map of the current time slot based on the policy and state of the current time slot to obtain the segmented second feature map of the current time slot, and send it to the vehicle; fuse the segmented second feature map of the current time slot with the first feature map of the current time slot through the feature fusion module to obtain the fused feature map of the current time slot; and input the fused feature map of the current time slot into the second feature decoding module to map the target category probability and target spatial location of the current time slot.

8. An electronic device, comprising: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.

Images