Target perception methods, systems, devices, and media
By acquiring perception data at multiple time points in the roadside perception system and performing feature extraction and encoding processing, the target perception result is predicted, solving the problem of excessive end-to-end latency and achieving faster and more accurate target detection, which is suitable for autonomous driving and intelligent traffic management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TIANYI TRANSPORTATION TECH CO LTD
- Filing Date
- 2025-06-27
- Publication Date
- 2026-07-24
Smart Images
Figure CN120748193B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of target detection, and more specifically to a target sensing method, system, device, and medium. Background Technology
[0002] A roadside perception system is a system installed beside a road to perceive the road environment and traffic conditions. Its common operating mode is as follows: first, various sensors equipped with the system (such as cameras, millimeter-wave radar, and lidar) collect data on the road. These sensors are responsible for collecting information such as the position, speed, and shape of traffic participants (including vehicles, pedestrians, and non-motorized vehicles), as well as relevant data on road conditions and the surrounding environment. After the sensors complete the data acquisition step, the system begins to process the collected data, such as using object detection algorithms to identify and detect traffic participants in images or point clouds.
[0003] End-to-end latency refers to the total time elapsed from when the sensor starts collecting data until the system outputs the final detection result. In this common operating mode of roadside perception systems, due to limitations of sensors and processing links, end-to-end latency can be quite long, even reaching 200 milliseconds (ms). For scenarios with high latency requirements (such as in autonomous driving scenarios where vehicles need to acquire real-time information about the surrounding environment to make rapid decisions to avoid collisions and ensure driving safety; or in intelligent traffic management where timely detection of traffic flow changes and corresponding adjustments to control strategies are required), a latency of 200 ms may lead to information lag, preventing the system from reacting accurately and promptly, thus failing to meet the needs of these scenarios and impacting system performance and safety. Summary of the Invention
[0004] In view of this, in order to overcome at least one aspect of the above problems, embodiments of the present invention propose a target perception method, comprising the following steps: Acquire the perception data at time T+M and time T+N; The query matrix at time T+M is obtained based on the sensing data at time T+M; The key matrix and value matrix at time T+N are obtained based on the sensing data at time T+N; The prediction result at time T+L is obtained using the query matrix at time T+M, the key matrix at time T+N, and the value matrix, and target perception is performed based on the prediction result.
[0005] In some embodiments, obtaining the perception result at time T+M based on the perception data at time T+M further includes: The target features at time T+M are obtained by using the feature extraction module to extract features from the perceived data at time T+M. The target features at time T+M are encoded using an encoding module to obtain the key matrix and value matrix at time T+M. Obtain the pre-trained first query matrix; The key matrix and value matrix at time T+M, as well as the first query matrix, are decoded using the first decoding module, and the decoding result is input into the first feedforward neural network to obtain the perception result at time T+M.
[0006] In some embodiments, it also includes: The feature extraction module, the encoding module, the first decoding module, and the first feedforward neural network are trained, and the object query module is trained to obtain the first query matrix.
[0007] In some embodiments, obtaining the key matrix and value matrix at time T+N based on the sensed data at time T+N further includes: The target features at time T+N are obtained by using the feature extraction module to extract features from the perceived data at time T+N. The target features at time T+N are encoded using an encoding module to obtain the key matrix and value matrix at time T+N.
[0008] In some embodiments, obtaining the prediction result at time T+L using the query matrix at time T+M, the key matrix at time T+N, and the value matrix further includes: The query matrix at time T+M, the key matrix at time T+N, and the value matrix are decoded using the second decoding module, and the decoding result is input into the second feedforward neural network to obtain the prediction result at time T+L.
[0009] In some embodiments, the prediction result at time T+L is obtained based on the query matrix at time T+M, the key matrix at time T+N, and the value matrix, and target perception is performed based on the prediction result, further including: The first prediction result at time T+L is obtained based on the query matrix at time T+X, the key matrix at time T+Y, and the value matrix, where X is less than Y; The second prediction result at time T+L is obtained based on the query matrix at time T+Y, the key matrix at time T+X, and the value matrix. A fusion result is obtained based on the first prediction result and the second prediction result, and target perception is performed based on the fusion result.
[0010] In some embodiments, obtaining a fusion result based on the first prediction result and the second prediction result, and performing target perception based on the fusion result, further includes: The target parameters and confidence scores of the fusion result are obtained based on the first target parameter and first confidence score in the first prediction result, and the second target parameter and second confidence score in the second prediction result; Target perception is performed based on the target parameters and confidence scores of the fusion results.
[0011] Based on the same inventive concept, according to another aspect of the present invention, embodiments of the present invention also provide a target perception system, comprising: The acquisition module is configured to acquire sensing data at time T+M and sensing data at time T+N; The first calculation module is configured to obtain the query matrix at time T+M based on the sensing data at time T+M; The second configuration module is configured to obtain the key matrix and value matrix at time T+N based on the sensing data at time T+N; The prediction module is configured to obtain the prediction result at time T+L using the query matrix at time T+M, the key matrix and the value matrix at time T+N, and to perform target perception based on the prediction result.
[0012] Based on the same inventive concept, according to another aspect of the present invention, embodiments of the present invention also provide a computer device, comprising: At least one processor; and A memory storing a computer program that can run on the processor, which, when executing the program, performs the steps of any of the target perception methods described above.
[0013] Based on the same inventive concept, according to another aspect of the present invention, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of any of the target perception methods described above.
[0014] The present invention has one of the following beneficial technical effects: Since traffic participants (such as vehicles, pedestrians, etc.) have inertia during their movement, their speed, position, heading angle (for objects with a direction of travel such as vehicles), and size (relatively fixed attributes such as the size of objects) will not change drastically in a short period of time. For example, a car driving normally will maintain a relatively stable speed, move its position according to a certain trajectory, and not change its direction of travel instantly if it does not brake suddenly, accelerate, or turn sharply. Therefore, the solution proposed in this invention predicts the perceived result at time T+L based on the perception result at time T+M and the encoding result corresponding to the perception data at time T+N, thereby significantly reducing the end-to-end latency of the roadside perception system from information collection to final processing and output of results. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.
[0016] Figure 1 A flowchart illustrating the target perception method provided in an embodiment of the present invention; Figure 2 A schematic diagram of the detection process of the target detection model provided for embodiments of the present invention; Figure 3 A flowchart illustrating a target perception method provided in an embodiment of the present invention; Figure 4 Another flowchart of the target perception method provided for embodiments of the present invention; Figure 5 This is a schematic diagram of the timeline for existing target detection. Figure 6 A schematic diagram of the timeline for target detection provided for embodiments of the present invention; Figure 7 A schematic diagram illustrating the fusion of perception and prediction results provided for an embodiment of the present invention; Figure 8 A schematic diagram of the target detection system provided for an embodiment of the present invention; Figure 9 A schematic diagram of the structure of a computer device provided for an embodiment of the present invention; Figure 10 A schematic diagram of the structure of a computer-readable storage medium provided for an embodiment of the present invention. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to specific examples and the accompanying drawings.
[0018] It should be noted that all uses of "first" and "second" in the embodiments of the present invention are for the purpose of distinguishing two entities or parameters with the same name but different names. It is clear that "first" and "second" are only for the convenience of expression and should not be construed as limiting the embodiments of the present invention. Subsequent embodiments will not explain this in detail.
[0019] According to one aspect of the present invention, embodiments of the present invention provide a target perception method, such as... Figure 1 As shown, it may include the following steps: S1, acquire the perception data at time T+M and time T+N; S2, Based on the sensing data at time T+M, obtain the query matrix at time T+M; S3, Based on the sensing data at time T+N, obtain the key matrix and value matrix at time T+N; S4. The prediction result at time T+L is obtained by using the query matrix at time T+M, the key matrix and the value matrix at time T+N, and target perception is performed based on the prediction result, where L is greater than M and L is greater than N.
[0020] Because traffic participants (such as vehicles and pedestrians) have inertia during their movement, their speed, position, heading angle (for vehicles and other objects with a direction of travel), and size (relatively fixed properties such as the size of an object) do not change drastically in a short period of time. For example, a car traveling normally will maintain a relatively stable speed, move along a certain trajectory, and not change its direction of travel instantly if it does not brake suddenly, accelerate, or turn sharply. Therefore, the solution proposed in this invention predicts the perceived result at time T+L based on the perception results at time T+M and the corresponding encoding results of the perception data at time T+N. This significantly reduces the end-to-end latency of the roadside perception system from information acquisition to final processing and output. At the same time, it utilizes the richer intermediate encoding features at time T+N for prediction, making the prediction more accurate.
[0021] In some embodiments, the sensing data may be data collected by a single sensor, such as data collected by lidar or millimeter-wave radar, or data collected by a camera, or a combination of data collected by multiple sensors. The sensing data is obtained by fusing the data through a multi-sensor data fusion network.
[0022] In some embodiments, obtaining the perception result at time T+M based on the perception data at time T+M further includes: The target features at time T+M are obtained by using the feature extraction module to extract features from the perceived data at time T+M. The target features at time T+M are encoded using an encoding module to obtain the key matrix and value matrix at time T+M. Obtain the pre-trained first query matrix; The key matrix and value matrix at time T+M, as well as the first query matrix, are decoded using the first decoding module, and the decoding result is input into the first feedforward neural network to obtain the perception result at time T+M.
[0023] Specifically, when obtaining the perception result at time T+M based on the perception data at time T+M, a target detection model can be used to detect the perception data at time T+M to obtain the perception result, such as... Figure 2 As shown, the target detection model can be a Transformer-based point cloud target detection model, such as TransFusion-L. It can include a feature extraction module, an encoding module, an object query module, a first decoding module, and a first feedforward neural network (first FFN). The object query module provides a first query matrix (Q). The feature extraction module can be a 3D Backbone module, used to extract features from the perceived data to obtain target features at time T+M, such as the target's BEV features. The encoding module encodes the target features to obtain the key matrix (K) and value matrix (V) at time T+M. After obtaining the first query matrix (Q), key matrix (K), and value matrix (V), the first decoding module decodes them. The decoding result is input into the first feedforward neural network to obtain the perceived result at time T+M, which can include the target category, position, size, and orientation. In some embodiments, M can be greater than N or less than N. Preferably, M is less than N, for example, N is 1, M is 0, and L is 2. In some embodiments, it also includes: The feature extraction module, the encoding module, the first decoding module, and the first feedforward neural network are trained, and the object query module is trained to obtain the first query matrix.
[0024] Specifically, a training set can be constructed, and the feature extraction module, encoding module, object query module, first decoding module, and first feedforward neural network can be trained using the training set. During training, gradient descent, as well as loss functions such as classification loss and location regression loss, can be used to train and optimize the model. The first query matrix can be obtained by training the object query module.
[0025] For example, using the formula Loss(GT,Det(F) Fused Training is performed where GT represents the ground truth, or the true value; Det represents the detection network, and F... Fused It is a multi-sensor fusion feature, namely F Fused =[BEV c BEV l BEV r [], where BEV represents the feature extraction module, and c, l, and r represent image data, LiDAR point cloud data, and millimeter-wave radar point cloud data, respectively. It should be noted that F Fused Adjustments are made based on the type of sensing data. If the sensing data consists only of LiDAR point cloud data, then F Fused =[ BEV l ].
[0026] In some embodiments, obtaining the key matrix and value matrix at time T+N based on the sensed data at time T+N further includes: The target features at time T+N are obtained by using the feature extraction module to extract features from the perceived data at time T+N. The target features at time T+N are encoded using an encoding module to obtain the key matrix and value matrix at time T+N.
[0027] Specifically, such as Figure 3 As shown, when predicting the perception results at time T+L, a target detection network is needed to process the perception data at times T+M and T+N. The perception data at time T+M requires the above process to obtain the perception results at time T+M. Then, the perception results at time T+M are used as the query matrix for time T+M, and the key matrix (K) and value matrix (V) are obtained based on the perception data at time T+N. Similarly, after acquiring the perception data at time T+N, the feature extraction module extracts target features, such as the BEV feature of the target, from the perception data at time T+N. Then, the encoding module encodes the target features to obtain the key matrix (K) and value matrix (V) at time T+N. Thus, after obtaining the query matrix, key matrix, and value matrix at time T+M, the perception results at time T+L can be predicted. This process eliminates the need to wait for the acquisition and processing at time T+L, reducing the end-to-end latency of the roadside perception system. Moreover, by utilizing the key matrix and value matrix at time T+N, i.e., intermediate features, the information contained is richer. When making predictions, combining the perception results at time T+M with the richer intermediate features at time T+N can yield more accurate prediction results.
[0028] For example, such as Figure 5 As shown, when N is 1, M is 0, and L is 2, according to the traditional target detection process, each frame of data needs to go through the processes of starting acquisition, acquisition completion, processing and transmission to obtain the final perception result, with an end-to-end latency of 200ms. Figure 6 As shown, according to the process proposed in this embodiment, the first and second frames are processed normally, and the prediction result of the third frame can be predicted after obtaining the perception result of the second frame. Therefore, starting from the third frame, the end-to-end latency can be reduced to 100ms.
[0029] In some embodiments, obtaining the prediction result at time T+L using the query matrix at time T+M, the key matrix at time T+N, and the value matrix further includes: The query matrix at time T+M, the key matrix at time T+N, and the value matrix are decoded using the second decoding module, and the decoding result is input into the second feedforward neural network to obtain the prediction result at time T+L.
[0030] Specifically, such as Figure 3 As shown, a second decoding module and a second feedforward neural network (second FFN) can be trained separately. Then, the trained second decoding module is used to decode and calculate the query matrix (Q) obtained based on the perception result at time T+M, as well as the key matrix (K) and value matrix (V) obtained based on time T+N. The calculation result is input into the second feedforward neural network to predict the perception result at time T+L.
[0031] When training the second decoding module and the second feedforward neural network, the loss function can be calculated based on the perception results at time T+L, or it can be calculated based on the ground truth values labeled at time T+L.
[0032] In some embodiments, the prediction result at time T+L is obtained based on the query matrix at time T+M, the key matrix at time T+N, and the value matrix, and target perception is performed based on the prediction result, further including: The first prediction result at time T+L is obtained based on the query matrix at time T+X, the key matrix at time T+Y, and the value matrix, where X is less than Y; The second prediction result at time T+L is obtained based on the query matrix at time T+Y, the key matrix at time T+X, and the value matrix. A fusion result is obtained based on the first prediction result and the second prediction result, and target perception is performed based on the fusion result.
[0033] Specifically, such as Figure 4As shown, similarly, the second decoding module decodes the query matrix (Q) obtained based on the perception results at time T+X, and the key matrix (K) and value matrix (V) obtained based on time T+Y. The calculation result is input into the second feedforward neural network to predict the first prediction result at time T+L. In some embodiments, a third decoding module and a third feedforward neural network (third FFN) can be trained separately. Then, the trained third decoding module decodes the query matrix (Q) obtained based on the perception results at time T+Y, and the key matrix (K) and value matrix (V) obtained based on time T+X. The calculation result is input into the third feedforward neural network (third FFN) to predict the second prediction result at time T+L. Finally, the first prediction result and the second prediction result are fused to obtain the final fused result. The first prediction result combines the rich intermediate coding features at time T+X, and the second prediction result combines the rich intermediate coding features at time T+Y. By fusing these two results, the advantages of both can be comprehensively utilized, making the final detection result more accurate and comprehensive. For example, in real-world scenarios, a target may be temporarily and dynamically occluded by other objects, leading to inaccuracies when using a single prediction result. However, by using two prediction results, the target's position and state at time T+L can still be predicted based on its previous motion trajectory and state, unaffected by the current occlusion. Thus, by fusing the two prediction results, missing information can be supplemented when the target is occluded, avoiding detection failures due to occlusion and improving the robustness of target detection.
[0034] In some embodiments, obtaining a fusion result based on the first prediction result and the second prediction result, and performing target perception based on the fusion result, further includes: The target parameters and confidence scores of the fusion result are obtained based on the first target parameter and first confidence score in the first prediction result, and the second target parameter and second confidence score in the second prediction result; Target perception is performed based on the target parameters and confidence scores of the fusion results.
[0035] Specifically, when fusing the first and second prediction results, the coordinates and confidence scores of the fused result can be obtained by taking the target parameters and confidence scores of both.
[0036] For example, such as Figure 7 As shown, BOX1 represents the first prediction result, X1, Y1, X1', and Y1' are the bounding box coordinates, and S1 is the confidence score. BOX2 represents the second prediction result, X2, Y2, X2', and Y2' are the bounding box coordinates, and S2 is the confidence score. The bounding box coordinates and confidence score of the fused result can be obtained based on the following formula:
[0037]
[0038]
[0039]
[0040]
[0041] Similarly, the target category, size, and orientation of the fusion result can all be determined using the formulas described above.
[0042] The proposed solution is based on the sensing results at time T+M and the encoding results corresponding to the sensing data at time T+N to predict the sensing results at time T+L, thereby significantly reducing the end-to-end latency of the roadside sensing system from information acquisition to final processing and output.
[0043] Based on the same inventive concept, according to another aspect of the present invention, embodiments of the present invention also provide a target perception system 400, such as... Figure 8 As shown, it includes: The acquisition module 401 is configured to acquire the sensing data at time T+M and the sensing data at time T+N; The first calculation module 402 is configured to obtain the query matrix at time T+M based on the sensing data at time T+M. The second configuration module 403 is configured to obtain the key matrix and value matrix at time T+N based on the sensing data at time T+N; The prediction module 404 is configured to obtain the prediction result at time T+L using the query matrix at time T+M, the key matrix and the value matrix at time T+N, and to perform target perception based on the prediction result, wherein M is greater than N.
[0044] In some embodiments, obtaining the perception result at time T+M based on the perception data at time T+M further includes: The target features at time T+M are obtained by using the feature extraction module to extract features from the perceived data at time T+M. The target features at time T+M are encoded using an encoding module to obtain the key matrix and value matrix at time T+M. Obtain the pre-trained first query matrix; The key matrix and value matrix at time T+M, as well as the first query matrix, are decoded using the first decoding module, and the decoding result is input into the first feedforward neural network to obtain the perception result at time T+M.
[0045] In some embodiments, it also includes: The feature extraction module, the encoding module, the first decoding module, and the first feedforward neural network are trained, and the object query module is trained to obtain the first query matrix.
[0046] In some embodiments, obtaining the key matrix and value matrix at time T+N based on the sensed data at time T+N further includes: The target features at time T+N are obtained by using the feature extraction module to extract features from the perceived data at time T+N. The target features at time T+N are encoded using an encoding module to obtain the key matrix and value matrix at time T+N.
[0047] In some embodiments, obtaining the prediction result at time T+L using the query matrix at time T+M, the key matrix at time T+N, and the value matrix further includes: The query matrix at time T+M, the key matrix at time T+N, and the value matrix are decoded using the second decoding module, and the decoding result is input into the second feedforward neural network to obtain the prediction result at time T+L.
[0048] In some embodiments, the prediction result at time T+L is obtained based on the query matrix at time T+M, the key matrix at time T+N, and the value matrix, and target perception is performed based on the prediction result, further including: The first prediction result at time T+L is obtained based on the query matrix at time T+X, the key matrix at time T+Y, and the value matrix, where X is less than Y; The second prediction result at time T+L is obtained based on the query matrix at time T+Y, the key matrix at time T+X, and the value matrix. A fusion result is obtained based on the first prediction result and the second prediction result, and target perception is performed based on the fusion result.
[0049] In some embodiments, obtaining a fusion result based on the first prediction result and the second prediction result, and performing target perception based on the fusion result, further includes: The target parameters and confidence scores of the fusion result are obtained based on the first target parameter and first confidence score in the first prediction result, and the second target parameter and second confidence score in the second prediction result; Target perception is performed based on the target parameters and confidence scores of the fusion results.
[0050] Based on the same inventive concept, according to another aspect of the present invention, such as Figure 9As shown, embodiments of the present invention also provide a computer device 501, comprising: At least one processor 520; and The memory 510 stores a computer program 511 that can run on a processor. When the processor 520 executes the program, it performs the steps of any of the target perception methods described above.
[0051] Based on the same inventive concept, according to another aspect of the present invention, such as Figure 10 As shown, embodiments of the present invention also provide a computer-readable storage medium 601, which stores a computer program 610. When the computer program 610 is executed by a processor, it performs the steps of any of the target perception methods described above.
[0052] Finally, it should be noted that those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods.
[0053] Furthermore, it should be understood that the computer-readable storage medium (e.g., memory) described herein may be volatile memory or non-volatile memory, or may include both volatile memory and non-volatile memory.
[0054] Those skilled in the art will also understand that the various exemplary logic blocks, modules, circuits, and algorithm steps described in conjunction with the disclosure herein can be implemented as electronic hardware, computer software, or a combination of both. To clearly illustrate this interchangeability between hardware and software, the functionality of various illustrative components, blocks, modules, circuits, and steps has been generally described. Whether this functionality is implemented as software or as hardware depends on the specific application and the design constraints imposed on the system as a whole. Those skilled in the art can implement the functionality in various ways for each specific application, but such implementation decisions should not be construed as departing from the scope of the embodiments disclosed herein.
[0055] The above are exemplary embodiments disclosed in this invention. However, it should be noted that various changes and modifications can be made without departing from the scope of the embodiments of this invention as defined by the claims. The functions, steps, and / or actions of the methods according to the disclosed embodiments described herein do not need to be performed in any particular order. Furthermore, although the elements disclosed in the embodiments of this invention may be described or claimed individually, they may be understood as multiple unless explicitly limited to a singular number.
[0056] It should be understood that, as used herein, the singular form “a” is intended to include the plural form as well, unless the context clearly supports an exception. It should also be understood that, as used herein, “and / or” refers to any and all possible combinations of one or more of the associated listed items.
[0057] The embodiment numbers disclosed in the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0058] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.
[0059] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples. Within the framework of the invention, technical features of the above embodiments or different embodiments can be combined, and many other variations of different aspects of the invention exist, which are not provided in the details for the sake of brevity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the protection scope of the invention.
Claims
1. A target perception method, characterized in that, Includes the following steps: Acquire the perception data at time T+M and time T+N; The query matrix at time T+M is obtained based on the sensing data at time T+M; The key matrix and value matrix at time T+N are obtained based on the sensing data at time T+N; Based on the query matrix at time T+M, the key matrix and value matrix at time T+N, the prediction result at time T+L is obtained, and target perception is performed based on the prediction result, where L is greater than M and L is greater than N. The query matrix at time T+M is obtained based on the sensing data at time T+M, further including: The target features at time T+M are obtained by extracting features from the perceived data at time T+M based on the feature extraction module. The key matrix and value matrix at time T+M are obtained by encoding the target features at time T+M using the encoding module. Obtain the pre-trained first query matrix; The first decoding module decodes the key matrix and value matrix at time T+M and the first query matrix, and inputs the decoding result into the first feedforward neural network to obtain the perception result at time T+M. The perception result at time T+M is then used as the query matrix at time T+M.
2. The method as described in claim 1, characterized in that, Also includes: The feature extraction module, the encoding module, the first decoding module, and the first feedforward neural network are trained, and the object query module is trained to obtain the first query matrix.
3. The method as described in claim 1, characterized in that, The key matrix and value matrix at time T+N are obtained based on the sensing data at time T+N, further including: The target features at time T+N are obtained by extracting features from the perceived data at time T+N based on the feature extraction module. The target features at time T+N are encoded using the encoding module to obtain the key matrix and value matrix at time T+N.
4. The method as described in claim 1, characterized in that, The prediction result at time T+L is obtained based on the query matrix at time T+M, the key matrix at time T+N, and the value matrix, further including: The second decoding module decodes the query matrix at time T+M, the key matrix at time T+N, and the value matrix, and inputs the decoding result into the second feedforward neural network to obtain the prediction result at time T+L.
5. The method as described in claim 1, characterized in that, Based on the query matrix at time T+M, the key matrix and value matrix at time T+N, the prediction result at time T+L is obtained, and target perception is performed based on the prediction result, further including: Based on the query matrix at time T+X, the key matrix at time T+Y, and the value matrix, the first prediction result at time T+L is obtained, where X is less than Y; The second prediction result at time T+L is obtained based on the query matrix at time T+Y, the key matrix at time T+X, and the value matrix. A fusion result is obtained based on the first prediction result and the second prediction result, and target perception is performed based on the fusion result.
6. The method as described in claim 5, characterized in that, A fusion result is obtained based on the first prediction result and the second prediction result. Target perception is then performed based on the fusion result, further including: The target parameters and confidence scores of the fusion result are obtained based on the first target parameter and first confidence score in the first prediction result, and the second target parameter and second confidence score in the second prediction result; Target perception is performed based on the target parameters and confidence scores of the fusion results.
7. A target perception system, characterized in that, include: The acquisition module is configured to acquire sensing data at time T+M and sensing data at time T+N; The first calculation module is configured to obtain the query matrix at time T+M based on the sensing data at time T+M; The second configuration module is configured to obtain the key matrix and value matrix at time T+N based on the sensing data at time T+N; The prediction module is configured to obtain the prediction result at time T+L using the query matrix at time T+M, the key matrix and the value matrix at time T+N, and to perform target perception based on the prediction result, where M is greater than N; The first computing module is also configured as follows: The target features at time T+M are obtained by extracting features from the perceived data at time T+M based on the feature extraction module. The key matrix and value matrix at time T+M are obtained by encoding the target features at time T+M using the encoding module. Obtain the pre-trained first query matrix; The first decoding module decodes the key matrix and value matrix at time T+M and the first query matrix, and inputs the decoding result into the first feedforward neural network to obtain the perception result at time T+M. The perception result at time T+M is then used as the query matrix at time T+M.
8. A computer device, comprising: At least one processor; as well as A memory storing a computer program executable on the processor, characterized in that the processor executes the program by performing the steps of the method as described in any one of claims 1-6.
9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it performs the steps of the method as described in any one of claims 1-6.