Memory enhancement based target detection method, system, device, and storage medium

Abstract

The application relates to the field of road anomaly analysis, and specifically discloses a target detection method, system and device based on memory enhancement and a storage medium, the method comprising the following steps: acquiring current frame data and performing feature coding to obtain first perception features; performing target detection based on the first perception features to obtain a current detection result comprising at least one current target; acquiring detection results of a plurality of historical targets in one or more previous frames; matching the historical detection result of each historical target with the first perception features; and fusing the historical detection result of the matched historical target with the current detection result and outputting a final detection result of the current frame. According to the scheme, the historical targets are stored and matched in subsequent frames, even if the target detection model misses detection, the targets can be recalled according to the memory principle, the detection recall rate of abnormal targets (such as a fallen bicycle, a lying pedestrian and the like) is effectively improved, and the safety of traffic is ensured.

Description

Technical Field

[0001] This invention relates to the field of target detection, and more specifically to a target detection method, system, device, and storage medium based on memory enhancement. Background Technology

[0002] Current roadside sensing systems still face serious issues with missed target detection in practical deployments. The sensors are installed relatively high (typically 6-8 meters) and far from the center of the road, resulting in distant targets occupying only a tiny pixel / point cloud area in the image or point cloud. Conventional detection models are already insufficient for detecting small targets. Even more challenging is the sharp increase in missed detection rates when targets undergo abnormal shape changes. Take a fallen bicycle as an example: an upright bicycle has clear structural features—two round wheels and a triangular frame—patterns that detection models focus on learning. However, after falling, its outline becomes a flat strip or a lump; the wheels are no longer round, and the frame geometry completely changes. Target detection models trained based on the "upright" pattern respond very weakly to these irregular shapes, leading to a sharp drop in detection rates. Furthermore, the contrast between the fallen target and the road background decreases, and environmental factors such as backlighting, shadows, rain, and fog further weaken weak signals, making missed detection particularly pronounced.

[0003] This missed detection issue has extremely serious consequences for traffic safety. For autonomous vehicles, a "fallen bicycle" is essentially a non-cooperative, irregularly shaped, static obstacle. Unlike vehicles or standing pedestrians, it is not easily identified as an object that path planning systems "must avoid." If the system misses the detection and makes a decision error, the vehicle may run it over directly, causing a serious accident. This type of scenario represents the highest level of protection needs for vulnerable road users. A bicycle falling over usually means an accident (such as a collision or skidding) or a physical condition of the rider (such as a sudden illness). In this situation, the rider completely loses the ability to move and avoid danger, making them extremely vulnerable to being run over by following vehicles, causing fatal secondary injuries. Therefore, the detection of objects such as fallen bicycles and lying pedestrians is crucial and requires a high recall rate. However, current road test perception systems are not specifically designed for this scenario and still use conventional object detection methods, occasionally resulting in missed detections and causing significant harm. 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 detection method based on memory enhancement, comprising the following steps: The first perceptual feature is obtained by acquiring the current frame data and performing feature encoding. Based on the first perceptual features, target detection is performed to obtain a current detection result that includes at least one current target; Obtain the detection results of multiple historical targets in one or more previous frames; The historical detection results of each historical target are matched with the first perceptual feature; The historical detection results of the successfully matched historical targets are fused with the current detection results to output the final detection result of the current frame.

[0005] In some embodiments, matching the historical detection results of each historical target with the first perceived feature further includes: Based on historical detection results, determine the coordinates and historical features corresponding to historical targets; Based on the historical coordinates, extract the real-time features of the corresponding location from the first perceived features; Calculate the similarity between the real-time features and the historical features; If the similarity is greater than a first threshold, a successful match is determined.

[0006] In some embodiments, the method further includes: The coordinates, category, and confidence level of the current target are determined based on the current detection results of the current target. Based on the coordinates of the current target, extract the corresponding features from the first perception feature; The coordinates, category, confidence level, and features of the current target are stored in a preset location as historical detection results corresponding to historical targets.

[0007] In some embodiments, the method further includes: The importance of each current target is determined based on the classifier; Multiple specific targets were selected based on the importance assessment results; The coordinates, category, confidence level, and features of the specific target are stored in a preset location as historical detection results corresponding to historical targets.

[0008] In some embodiments, the method further includes: Calculate the similarity between the features of the current target or a specific target and the features of each stored historical target; If the similarity is greater than the second threshold, it is determined to be a duplicate target. The storage timestamp of the existing historical target is abandoned or updated. When the number of stored historical targets exceeds the preset limit or the storage time exceeds the preset time window, the earliest stored historical target is removed first.

[0009] In some embodiments, fusing historical detection results of successfully matched historical targets with the current detection results and outputting the final detection result of the current frame further includes: Based on nonmaximum suppression, the historical detection results of the successfully matched historical targets are fused with the current detection results.

[0010] In some embodiments, acquiring current frame data and performing feature encoding to obtain first perceptual features further includes: Acquire at least one modal data from the following: lidar point cloud data, camera image data, and millimeter-wave radar data collected by sensors; Based on the fixed installation position of the sensor, a region of interest is preset, and the modal data is spatially cropped to retain the effective data within the region of interest; The first perceptual feature is obtained by feature encoding based on the effective data.

[0011] Based on the same inventive concept, according to another aspect of the present invention, embodiments of the present invention also provide a memory-enhanced target detection system, comprising: The first acquisition module is configured to acquire the current frame data and perform feature encoding to obtain the first perceptual feature; The target detection module is configured to perform target detection based on the first perception feature to obtain a current detection result including at least one current target; The second acquisition module is configured to acquire the detection results of multiple historical targets in one or more previous frames; The matching module is configured to match the historical detection results of each historical target with the first perceived feature; The fusion module is configured to fuse the historical detection results of the successfully matched historical target with the current detection results and output the final detection result of the current frame.

[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 The memory stores a computer program that can run on the processor, which, when executing the program, performs the steps of any of the memory-enhanced target detection 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 memory-enhanced target detection methods described above.

[0014] The present invention has one of the following beneficial technical effects: The solution proposed in this invention stores historical targets and performs matching recall in subsequent frames. Even if the target detection model misses a target, it can recall the target according to the memory principle, which effectively improves the detection recall rate of abnormal targets (such as fallen bicycles, lying pedestrians, etc.) and ensures the safety of autonomous driving and intelligent transportation. 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 schematic flowchart of a memory-enhanced target detection method provided for an embodiment of the present invention; Figure 2 A schematic diagram of a memory-enhanced target detection system provided for an embodiment of the present invention; Figure 3 A schematic diagram of the structure of a computer device provided for an embodiment of the present invention; Figure 4 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 propose a target detection method based on memory enhancement, such as... Figure 1 As shown, it may include the following steps: S1, acquire the current frame data and perform feature encoding to obtain the first perceptual feature; S2, based on the first perceptual features, target detection is performed to obtain a current detection result including at least one current target; S3, obtain the detection results of multiple historical targets in one or more previous frames; S4, match the historical detection results of each historical target with the first perceived feature; S5, the historical detection results of the successfully matched historical targets are fused with the current detection results, and the final detection result of the current frame is output.

[0020] Specifically, the process begins by acquiring the current frame data from a sensor, which can be one or more of a LiDAR, camera, or millimeter-wave radar. Then, feature encoding is performed on the current frame data to extract the first perceptual features. For LiDAR point cloud data, BEV (bird's eye view) features can be extracted. Specifically, a Transformer-based point cloud object detection algorithm (such as TransFusion-L) can be used to learn and train modules such as 3D Backbone, ObjectQueries, Object Detection Transformer Decoder, and Object Detection FFN to extract BEV features. F BEV .

[0021] After obtaining the first perceptual features, a pre-trained object detection model is used to detect objects, yielding the detection results for the current frame. The object detection model outputs information including the object's category, location (bounding box coordinates), size, orientation, and confidence score. This object detection model can be trained using labeled 3D object information (including object category, location, size, and orientation) and a common gradient descent optimization method. The loss function includes classification loss and location regression loss: Loss(GT, Det( F BEV In this context, GT represents the ground truth, and Det represents the detection network. F BEV This represents the extracted BEV features.

[0022] Next, the detection results of multiple pre-stored historical targets are retrieved from the history memory. These historical target detection results can include the target's coordinates, category, confidence level, and feature information. Then, the detection result of each historical target is matched with the first perceptual feature of the current frame, and based on the stored coordinates, the detection results of the target in the current frame are retrieved. F BEV Features are selected from the data, and the similarity between the historical target and the features at the same position in the current frame is compared to determine whether the historical target still exists in the current frame.

[0023] Finally, the historical detection results of successfully matched historical targets are fused with the current detection results to output the final detection result of the current frame.

[0024] The proposed solution stores historical targets and performs matching and recall in subsequent frames. Even if the target detection model misses a target, it can still recall the target based on the memory principle, which effectively improves the detection recall rate of abnormal targets (such as fallen bicycles, lying pedestrians, etc.) and ensures the safety of autonomous driving and intelligent transportation.

[0025] In some embodiments, matching the historical detection results of each historical target with the first perceived feature further includes: Based on historical detection results, determine the coordinates and historical features corresponding to historical targets; Based on the historical coordinates, extract the real-time features of the corresponding location from the first perceived features; Calculate the similarity between the real-time features and the historical features; If the similarity is greater than a first threshold, a successful match is determined.

[0026] Specifically, for each historical target, the detection result of that historical target is first read from the historical memory to determine its corresponding coordinates (bounding box) and historical features. The historical features can be feature vectors extracted and saved based on the BEV features of the frame in which the historical target was stored.

[0027] Then, using these historical coordinates, features corresponding to the given coordinates are extracted from the first perceptual features of the current frame, serving as real-time features. Next, the similarity between these real-time features and historical features is calculated. Similarity can be calculated using common methods such as cosine similarity and mutual information. Sim( F BEV [Box History ], F History [Box History ]) Sim represents similarity calculation. F BEV [Box History This indicates that the real-time features are extracted from the first perceptual features of the current frame using the coordinates of historical targets. F History [Box History [This is a historical characteristic.]

[0028] A first threshold, i.e., a similarity threshold, can be preset, and the calculated similarity is compared with this threshold. If the similarity is greater than the first threshold, it means that the historical target still exists in the same position in the current frame. Even if the target detection model in the current frame does not detect the target, the target can be recalled according to the memory principle; otherwise, it is determined that the historical target has disappeared or moved in the current frame, and the match fails.

[0029] This method, by extracting real-time features from the current frame based on coordinates and comparing them with historical features, accurately determines whether a historical target still exists in the current frame. Even if only part of the target is visible (e.g., only the rear of a vehicle), the missed targets can still be recalled because the historical detection stores the features of the complete target (e.g., the entire vehicle). This fully utilizes the fixed location of the roadside sensors and achieves efficient target matching.

[0030] In some embodiments, the method further includes: The coordinates, category, and confidence level of the current target are determined based on the current detection results of the current target. Based on the coordinates of the current target, extract the corresponding features from the first perception feature; The coordinates, category, confidence level, and features of the current target are stored in a preset location as historical detection results corresponding to historical targets.

[0031] Specifically, after obtaining the current detection result, for each current target in the current detection result, the coordinates (bounding box), category, and confidence score of the target are first extracted from the detection result. Then, using the coordinates of the target, features corresponding to the coordinates are extracted from the first perceptual feature map. Specifically, the coordinates of the target detected in the current frame are used as an index to select features of the corresponding region from the BEV feature map.

[0032] Then, the target's coordinates, category, confidence level, and extracted features are associated and stored as a complete historical detection result in a preset location in the historical storage. During storage, the storage is managed according to the first-in, first-out principle, and the physical length or time length can be set as needed, for example, storing 100 targets or storing all targets within 1 minute.

[0033] In some embodiments, the method further includes: The importance of each current target is determined based on the classifier; Multiple specific targets were selected based on the importance assessment results; The coordinates, category, confidence level, and features of the specific target are stored in a preset location as historical detection results corresponding to historical targets.

[0034] Specifically, before storing historical targets, an importance judgment mechanism can be set up, that is, a pre-trained classifier can be set up to determine whether the detected targets belong to important targets that need to be focused on. This classifier can be a binary classifier (distinguishing between important targets and unimportant targets), or multiple binary classifiers (each classifying the importance of different categories, such as important bicycles versus unimportant bicycles, important pedestrians versus unimportant pedestrians, etc.).

[0035] The classifier's input is features selected from the BEV (Browser Impact Value) features, based on the coordinates of the detected target in the current frame. It primarily uses convolution and fully connected layers to process these input features, outputting the probability that the target belongs to an important target category. The classifier is trained using the common gradient descent optimization method, and the loss function can be a classification loss. Loss (GT, C ( F BEV [Box])) Where GT represents the ground truth label of whether the target is an important target, i.e., whether it is an important target; C represents the classifier; F BEV [Box] indicates that the coordinates of the current target are used to navigate from F. BEV Features extracted from [the data].

[0036] Finally, based on the classifier's judgment, specific targets deemed important are selected from all current targets, such as bicycles about to fall or pedestrians about to trip—high-risk targets. The coordinates, category, confidence level, and extracted features of these specific targets are then saved to a preset location in the history storage. Storage follows a first-in, first-out (FIFO) principle, and the physical or time length can be set as needed, for example, storing 100 targets or all targets within one minute.

[0037] By setting a classifier to filter targets based on their importance, only important targets that require focused attention are stored. This avoids the memory waste caused by storing a large number of irrelevant targets, improves storage efficiency and subsequent matching computation efficiency, and enables the system to focus more on high-risk scenarios, thus better protecting the safety of vulnerable traffic participants.

[0038] In some embodiments, the method further includes: Calculate the similarity between the features of the current target or a specific target and the features of each stored historical target; If the similarity is greater than the second threshold, it is determined to be a duplicate target. The storage timestamp of the existing historical target is abandoned or updated. When the number of stored historical targets exceeds the preset limit or the storage time exceeds the preset time window, the earliest stored historical target is removed first.

[0039] Specifically, before storing the current target (e.g., all targets detected in this test) or a specific target into the history memory, the similarity between the features of the target and the features of each historical target already stored in the history memory is first calculated. The similarity can be calculated using common similarity calculation methods such as cosine similarity and mutual information.

[0040] The calculated similarity is compared to a preset second threshold, also known as a deduplication threshold. If the similarity is greater than the deduplication threshold, it means that the current target and a previously stored historical target are the same target, and the target is identified as a duplicate. In this case, the target can be discarded to avoid duplicate records in memory, or the storage timestamp of an existing historical target can be updated to extend its retention time in memory. If the similarity is not greater than the second threshold, the target is identified as a new target and stored in the historical memory.

[0041] In terms of storage space management, the historical storage can be managed according to the first-in, first-out (FIFO) principle. When the number of stored historical targets exceeds the preset limit (e.g., 100 targets) or the storage time exceeds the preset time window (e.g., 1 minute), the earliest stored historical target is removed first to free up storage space for new important targets.

[0042] This deduplication mechanism avoids duplicate storage of the same target in memory, saving storage resources. Simultaneously, the first-in, first-out (FIFO) memory management strategy ensures that the memory always holds the latest information on important targets, preventing memory overflow.

[0043] In some embodiments, fusing historical detection results of successfully matched historical targets with the current detection results and outputting the final detection result of the current frame further includes: Based on nonmaximum suppression, the historical detection results of the successfully matched historical targets are fused with the current detection results.

[0044] Specifically, after matching the historical target with the first perceptual feature of the current frame, the historical detection results of the successfully matched historical targets are obtained, i.e., targets recalled according to the memory principle. Then, these historical detection results are fused together with the current detection results.

[0045] The fusion method can employ the NMS (Non-Maximum Suppression) algorithm. For the same target detected in both the current detection result and the successfully matched historical detection results, the NMS algorithm retains the result with higher detection confidence and removes redundant detection boxes. For targets not detected in the current detection result but successfully matched in historical detection results (i.e., targets missed in the current frame but recalled based on memory), these historical detection results are used as valid detection outputs in the NMS fusion and are ultimately retained in the output result.

[0046] By using NMS fusion, the final detection result for the current frame is obtained, which includes both targets correctly detected by the current frame detection model and targets missed by the model based on memory principles. This organically integrates the recalled targets with the current detection result, eliminating redundancy from duplicate detections and ensuring that missed targets are successfully recalled and output, thus improving the completeness and accuracy of the detection results.

[0047] In some embodiments, acquiring current frame data and performing feature encoding to obtain first perceptual features further includes: Acquire at least one modal data from the following: lidar point cloud data, camera image data, and millimeter-wave radar data collected by sensors; Based on the fixed installation position of the sensor, a region of interest is preset, and the modal data is spatially cropped to retain the effective data within the region of interest; The first perceptual feature is obtained by feature encoding based on the effective data.

[0048] Specifically, the present invention can support input from multiple sensor modalities, including at least one of lidar point cloud data, camera image data, and millimeter-wave radar data. It can also support combined input of multiple sensor data. Multimodal fusion can further improve the accuracy and robustness of detection.

[0049] Since roadside sensors are fixedly installed and their observation range is fixed, a Region of Interest (ROI) can be pre-defined, and only data within the ROI can be processed. The ROI can be set manually, for example, by manually identifying areas highly related to weather (such as the sky) and removing areas unrelated to weather or with little correlation, such as buildings; alternatively, the ROI can be automatically generated using semantic segmentation models (such as Mask2Former). c roi =ROI(c) Where c represents the original data, c roi This represents the valid data retained within the ROI after cropping, utilizing only c. roiThe relevant data is used to train the model and perform inference.

[0050] Then, feature encoding is performed on the effective data within the cropped ROI to obtain the first sensing feature. When the input includes only one type of sensor data, the first sensing feature can be obtained by directly encoding the effective data of that modality. When the input includes multiple types of sensor data, modality-specific features are extracted from the effective data of each modality, and then the features of each modality are fused and encoded to obtain the first sensing feature. For example, the features of each modality can be uniformly mapped to the BEV space for fusion.

[0051] Based on the same inventive concept, according to another aspect of the present invention, embodiments of the present invention also provide a memory-enhanced target detection system 200, such as... Figure 2 As shown, it includes: The first acquisition module 201 is configured to acquire the current frame data and perform feature encoding to obtain the first perceptual feature; The target detection module 202 is configured to perform target detection based on the first perception features to obtain a current detection result including at least one current target; The second acquisition module 203 is configured to acquire the detection results of multiple historical targets in one or more previous frames; The matching module 204 is configured to match the historical detection results of each of the historical targets with the first perceptual features; The fusion module 205 is configured to fuse the historical detection results of the successfully matched historical target with the current detection results and output the final detection result of the current frame.

[0052] In some embodiments, matching the historical detection results of each historical target with the first perceived feature further includes: Based on historical detection results, determine the coordinates and historical features corresponding to historical targets; Based on the historical coordinates, extract the real-time features of the corresponding location from the first perceived features; Calculate the similarity between the real-time features and the historical features; If the similarity is greater than a first threshold, a successful match is determined.

[0053] In some embodiments, a storage module is also included, configured as follows: The coordinates, category, and confidence level of the current target are determined based on the current detection results of the current target. Based on the coordinates of the current target, extract the corresponding features from the first perception feature; The coordinates, category, confidence level, and features of the current target are stored in a preset location as historical detection results corresponding to historical targets.

[0054] In some embodiments, the storage module is further configured as follows: The importance of each current target is determined based on the classifier; Multiple specific targets were selected based on the importance assessment results; The coordinates, category, confidence level, and features of the specific target are stored in a preset location as historical detection results corresponding to historical targets.

[0055] In some embodiments, the storage module is further configured as follows: Calculate the similarity between the features of the current target or a specific target and the features of each stored historical target; If the similarity is greater than the second threshold, it is determined to be a duplicate target. The storage timestamp of the existing historical target is abandoned or updated. When the number of stored historical targets exceeds the preset limit or the storage time exceeds the preset time window, the earliest stored historical target is removed first.

[0056] In some embodiments, fusing historical detection results of successfully matched historical targets with the current detection results and outputting the final detection result of the current frame further includes: Based on nonmaximum suppression, the historical detection results of the successfully matched historical targets are fused with the current detection results.

[0057] In some embodiments, acquiring current frame data and performing feature encoding to obtain first perceptual features further includes: Acquire at least one modal data from the following: lidar point cloud data, camera image data, and millimeter-wave radar data collected by sensors; Based on the fixed installation position of the sensor, a region of interest is preset, and the modal data is spatially cropped to retain the effective data within the region of interest; The first perceptual feature is obtained by feature encoding based on the effective data.

[0058] Based on the same inventive concept, according to another aspect of the present invention, such as Figure 3 As 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 memory-enhanced target detection methods described above.

[0059] Based on the same inventive concept, according to another aspect of the present invention, such as Figure 4As 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 memory-enhanced target detection methods described above.

[0060] 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.

[0061] 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.

[0062] 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, 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.

[0063] 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.

[0064] 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.

[0065] 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.

[0066] 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.

[0067] 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 detection method based on memory enhancement, characterized in that, Includes the following steps: The first perceptual feature is obtained by acquiring the current frame data and performing feature encoding. Based on the first perceptual features, target detection is performed to obtain a current detection result that includes at least one current target; Obtain the detection results of multiple historical targets in one or more previous frames; The historical detection results of each historical target are matched with the first perceptual feature; The historical detection results of the successfully matched historical targets are fused with the current detection results to output the final detection result of the current frame.

2. The method as described in claim 1, characterized in that, Matching the historical detection results of each historical target with the first perceived feature further includes: Based on historical detection results, determine the coordinates and historical features corresponding to historical targets; Based on the historical coordinates, extract the real-time features of the corresponding location from the first perceived features; Calculate the similarity between the real-time features and the historical features; If the similarity is greater than a first threshold, a successful match is determined.

3. The method as described in claim 1, characterized in that, Also includes: The coordinates, category, and confidence level of the current target are determined based on the current detection results of the current target. Based on the coordinates of the current target, extract the corresponding features from the first perception feature; The coordinates, category, confidence level, and features of the current target are stored in a preset location as historical detection results corresponding to historical targets.

4. The method as described in claim 3, characterized in that, Also includes: The importance of each current target is determined based on the classifier; Multiple specific targets were selected based on the importance assessment results; The coordinates, category, confidence level, and features of the specific target are stored in a preset location as historical detection results corresponding to historical targets.

5. The method as described in claim 3 or 4, characterized in that, Also includes: Calculate the similarity between the features of the current target or a specific target and the features of each stored historical target; If the similarity is greater than the second threshold, it is determined to be a duplicate target. The storage of the existing historical target is abandoned or the storage timestamp of the existing historical target is updated. When the number of stored historical targets exceeds the preset limit or the storage time exceeds the preset time window, the earliest stored historical target is removed first.

6. The method as described in claim 1, characterized in that, The final detection result of the current frame is output by fusing the historical detection results of the successfully matched historical targets with the current detection results, and further includes: Based on nonmaximum suppression, the historical detection results of the successfully matched historical targets are fused with the current detection results.

7. The method as described in claim 1, characterized in that, The first perceptual feature is obtained by acquiring the current frame data and performing feature encoding, which further includes: Acquire at least one modal data from the following: lidar point cloud data, camera image data, and millimeter-wave radar data collected by sensors; Based on the fixed installation position of the sensor, a region of interest is preset, and the modal data is spatially cropped to retain the effective data within the region of interest; The first perceptual feature is obtained by feature encoding based on the effective data.

8. A target detection system based on memory enhancement, characterized in that, include: The first acquisition module is configured to acquire the current frame data and perform feature encoding to obtain the first perceptual feature; The target detection module is configured to perform target detection based on the first perception feature to obtain a current detection result including at least one current target; The second acquisition module is configured to acquire the detection results of multiple historical targets in one or more previous frames; The matching module is configured to match the historical detection results of each historical target with the first perceived feature; The fusion module is configured to fuse the historical detection results of the successfully matched historical target with the current detection results and output the final detection result of the current frame.

9. 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 1-7.

10. 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 1-7.

Images