Target detection method and device with perspective supervision, equipment and storage medium

Abstract

The invention discloses a target detection method and device with perspective supervision, equipment and a storage medium, and relates to the technical field of automatic driving, and the method comprises the steps: obtaining camera parameters, multi-view image information, a historical combination instance, a first combination instance and a second combination instance, and enabling the combination instance to comprise an anchor point and a feature corresponding to the anchor point; processing according to the camera parameters, the multi-view image information, the historical combination instance, the first combination instance and the second combination instance to obtain a target combination instance; and decoding the target combination instance to realize target detection, so that the problem of how to improve the flexibility and accuracy of target detection is solved, and the detection performance and the reasoning speed are improved.

Description

Technical Field

[0001] The present application relates to the technical field of autonomous driving, and in particular to a target detection method, device, equipment and storage medium with perspective supervision. Background Art

[0002] In autonomous driving technology, target detection is one of the core tasks of the perception system, and target detection under BEV (Bird's Eye View) has attracted much attention because it can provide more comprehensive and intuitive traffic scene information. At present, there are two main methods for achieving BEV target detection. One method is to stitch images taken by multiple cameras into a bird's-eye view through inverse perspective transformation, and then use a deep learning network for target detection. This method often has image stitching errors at the intersection of FOV (Field of View), resulting in image quality degradation and affecting detection performance. At the same time, objects above the ground will produce erroneous projections during the inverse perspective transformation process, making the detection results inconsistent with the actual situation. Another method directly detects 3D targets in the perspective image, and then generates 3D targets under the BEV perspective through perspective transformation. Although this method avoids the problem of image stitching, it requires the fusion of targets detected by multiple cameras, has high requirements for target tracking across lenses, and has high model complexity and high computing resource consumption.

[0003] Therefore, how to improve the flexibility and accuracy of target detection is a problem that needs to be solved urgently. Summary of the invention

[0004] The main purpose of this application is to provide a target detection method, device, equipment and storage medium with perspective supervision, aiming to solve the technical problem of how to improve the flexibility and accuracy of target detection.

[0005] To achieve the above objectives, the present application proposes a target detection method with perspective supervision, the method comprising:

[0006] Obtaining camera parameters, multi-view image information, a historical combination instance, a first combination instance, and a second combination instance, wherein the combination instance includes an anchor point and features corresponding thereto;

[0007] Processing is performed according to the camera parameters, the multi-view image information, the historical combination instance, the first combination instance, and the second combination instance to obtain a target combination instance;

[0008] The target combination instance is decoded to achieve target detection.

[0009] In one embodiment, the step of obtaining a target combination instance by processing according to the camera parameters, the multi-view image information, the historical combination instance, the first combination instance, and the second combination instance comprises:

[0010] Obtain multi-view image features, detection results, and perspective 3D target anchor points according to the camera parameters, the multi-view image information, and the first combination instance;

[0011] A target combination instance is obtained by processing according to the detection result, the historical combination instance, the second combination instance, the multi-view image feature and the perspective 3D target anchor point.

[0012] In one embodiment, the step of obtaining multi-view image features, detection results, and perspective 3D target anchor points according to the camera parameters, the multi-view image information, and the first combination instance includes:

[0013] Extracting image features from the multi-view images to obtain multi-view image features;

[0014] Supervise the image feature extraction process to obtain perspective 3D target anchor points;

[0015] A detection result is obtained according to the multi-view image features, the camera parameters and the first combination instance.

[0016] In one embodiment, the step of supervising the image feature extraction process to obtain the perspective 3D object anchor point includes:

[0017] Supervising the image feature extraction process according to a first preset supervision method to obtain an initial perspective 3D target anchor point;

[0018] Supervising the image feature extraction process according to a second preset supervision method to obtain depth information;

[0019] A perspective 3D object anchor point is obtained according to the initial perspective 3D object anchor point and the depth information.

[0020] In one embodiment, the step of obtaining the detection result according to the multi-view image features, the camera parameters and the first combination instance includes:

[0021] Mapping the camera parameters according to a preset position conversion strategy to obtain a high-dimensional feature vector;

[0022] Performing feature combination according to the high-dimensional feature vector and the first combination instance to obtain a third combination instance;

[0023] According to a preset sampling strategy, feature sampling and weighted summation are performed on the camera parameters, the third combination instance, and the multi-view image features to obtain a detection result.

[0024] In one embodiment, the step of obtaining a historical combination instance includes:

[0025] Obtain the vehicle's motion posture and initial history combination instance;

[0026] The initial historical combination instance is coordinate-converted according to the vehicle motion posture to obtain a historical combination instance.

[0027] In one embodiment, the step of obtaining the first combination instance and the second combination instance includes:

[0028] Obtain detection space and real annotation data;

[0029] Initializing the detection space according to a preset mean strategy to obtain a first combination instance;

[0030] Noise is added according to the real labeled data to obtain a second combination instance.

[0031] In addition, to achieve the above objectives, the present application also proposes a target detection device with perspective supervision, the device comprising:

[0032] A data acquisition module, used to acquire camera parameters, multi-view image information, historical combination instances, first combination instances, and second combination instances, where the combination instances include anchor points and their corresponding features;

[0033] A data processing module, configured to obtain a target combination instance by processing according to the camera parameters, the multi-view image information, the historical combination instance, the first combination instance, and the second combination instance;

[0034] The target detection module is used to decode the target combination instance to achieve target detection.

[0035] In addition, to achieve the above-mentioned objectives, the present application also proposes a target detection device with perspective supervision, the device comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, the computer program being configured to implement the steps of the target detection method with perspective supervision as described above.

[0036] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the target detection method with perspective supervision as described above are implemented.

[0037] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the steps of the target detection method with perspective supervision as described above.

[0038] The present application provides a target detection method with perspective supervision, and the method of the present application includes: obtaining camera parameters, multi-view image information, historical combination instances, first combination instances and second combination instances, the combination instance including anchor points and their corresponding features; processing according to the camera parameters, the multi-view image information, the historical combination instances, the first combination instances and the second combination instances to obtain target combination instances; decoding the target combination instances to achieve target detection. In summary, the present application improves the adaptability of the image backbone to 3D target detection tasks through perspective view supervision, and adopts an anchor-based sparse BEV detector scheme, which is more flexible in temporal fusion, solves the problem of how to improve the flexibility and accuracy of target detection, and improves detection performance and reasoning speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0040] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0041] Figure 1 A flowchart diagram of the first embodiment of the object detection method with perspective supervision provided in the present application;

[0042] Figure 2 This is a schematic diagram of a network structure in an embodiment of a target detection method with perspective supervision of the present application;

[0043] Figure 3 A schematic diagram of decoding of a 3D detection example in an embodiment of an object detection method with perspective supervision of the present application;

[0044] Figure 4 A flowchart diagram of a second embodiment of the object detection method with perspective supervision provided in the present application;

[0045] Figure 5 This is a schematic diagram of the module structure of the target detection device with perspective supervision according to an embodiment of the present application;

[0046] Figure 6Schematic diagram of the device structure of the hardware operating environment involved in the target detection method with perspective supervision in the embodiment of the present application.

[0047] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0048] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0049] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0050] The main solution of the embodiment of the present application is: obtaining camera parameters, multi-view image information, historical combination instances, first combination instances and second combination instances, where the combination instance includes an anchor point and its corresponding features; processing according to the camera parameters, the multi-view image information, the historical combination instance, the first combination instance and the second combination instance to obtain a target combination instance; decoding the target combination instance to achieve target detection.

[0051] In autonomous driving technology, target detection is one of the core tasks of the perception system, and target detection under BEV has attracted much attention because it can provide more comprehensive and intuitive traffic scene information. At present, the methods for achieving BEV target detection are mainly divided into two categories. One method is to stitch images taken by multiple cameras into a bird's-eye view through inverse perspective transformation, and then use a deep learning network for target detection. This method often has image stitching errors at the intersection of FOVs, resulting in image quality degradation and affecting detection performance. At the same time, objects above the ground will produce incorrect projections during the inverse perspective transformation process, making the detection results inconsistent with the actual situation. Another method directly detects 3D targets in the perspective image and then generates 3D targets under the BEV perspective through perspective transformation. Although this method avoids the problem of image stitching, it requires the fusion of targets detected by multiple cameras, has high requirements for target tracking across lenses, and has high model complexity and large consumption of computing resources. Therefore, how to improve the flexibility and accuracy of target detection is a problem that needs to be solved urgently.

[0052] This application improves the adaptability of the image backbone to 3D target detection tasks through perspective view supervision, and adopts an anchor-based sparse BEV detector scheme, which is more flexible in temporal fusion. It solves the problem of how to improve the flexibility and accuracy of target detection, and improves detection performance and reasoning speed.

[0053] It should be noted that the execution subject of this embodiment can be a target detection system with perspective supervision, or a computing service device with data processing, network communication and program running functions, or an electronic device capable of realizing the above-mentioned target detection function with perspective supervision, etc., and this embodiment does not specifically limit this. The following takes the target detection system with perspective supervision as an example to illustrate this embodiment and the following embodiments.

[0054] Based on this, the present application embodiment provides a target detection method with perspective supervision, referring to Figure 1 , Figure 1 Schematic diagram of the flow chart of the first embodiment of the object detection method with perspective supervision of the present application.

[0055] In this embodiment, the target detection method with perspective supervision includes steps S10 to S30:

[0056] Step S10: Obtain camera parameters, multi-view image information, historical combination instances, first combination instances, and second combination instances, where the combination instances include anchor points and their corresponding features.

[0057] It should be noted that in this step, the system will extract real-time image information, that is, multi-view image information, through multiple cameras installed at different positions on the vehicle. Camera parameters include internal and external parameters of the camera, such as focal length, optical center position, rotation matrix and translation vector, etc. The first combination instance and the second combination instance are two different initialization instances, and there is no order relationship between the first and the second. In addition, it should be noted that the anchor point in the combination instance not only contains the location information of the target, but also contains prior information such as the shape and size of the target. Features are high-order semantic descriptions of the target, which help to distinguish targets of different categories. Historical combination instances refer to target instances detected in the previous frame or the previous frames, including anchor points (prior information of the bounding box) and corresponding features (high-order semantic features extracted from the image by the image encoder).

[0058] It can be understood that this embodiment adopts the anchor-based sparse BEV detector. Due to the prior information of the anchor point, the anchor+feature instance is decoupled, making the sparse detection more flexible in time series fusion.

[0059] In a feasible implementation manner, the step of obtaining a historical combination instance includes:

[0060] Step A10: Obtain the vehicle's motion posture and initial history combination instance.

[0061] It should be noted that the vehicle motion posture refers to the dynamic information such as the position, direction, speed, etc. of the vehicle in three-dimensional space. The initial historical combination instance refers to the set of target detection results in the previous frame that have not been converted, including the predicted position of the target and the feature information extracted from the image, which is used to assist the detection of the current frame. In addition, it should be noted that in this step, the system will obtain the vehicle's motion posture information in real time through on-board sensors (such as GPS, gyroscope, etc.), including position, direction, speed, etc. Next, the system extracts the combination instances of the most recent frames (the specific number of frames depends on the system design) from the historical database.

[0062] Step A20: performing coordinate transformation on the initial historical combination instance according to the vehicle motion posture to obtain a historical combination instance.

[0063] It should be noted that if Figure 2 As shown, in this step, the system will perform coordinate transformation on the anchor points in the initial historical combination instance (i.e., T-1) according to the acquired vehicle motion posture information, and transform them from their respective historical coordinate systems to the coordinate system of the current frame. It can be understood that through coordinate transformation, the reuse of historical data in the current frame is realized, so that the detector can utilize the information in the time series, improving the accuracy and stability of target detection.

[0064] In addition, it should be noted that during the coordinate transformation process, the feature part (high-order semantic information) in the combined instance remains unchanged, because these features describe the inherent properties of the target and do not change with the movement of the vehicle.

[0065] In a feasible implementation manner, the step of obtaining the first combination instance and the second combination instance includes:

[0066] Step A30: Obtain detection space and real annotation data.

[0067] It should be noted that in this step, the system will define a three-dimensional detection space according to the perception requirements of the pre-designed autonomous driving vehicle. At the same time, the system will use existing data sets or collect data in the environment in real time through sensors (such as cameras, LiDAR), and annotate the data manually or automatically to obtain real annotated data containing information such as target position, size, speed, etc. In addition, it should be noted that the size and shape of the detection space can be adjusted according to the actual application scenario to ensure the comprehensiveness and accuracy of target detection.

[0068] Step A40: Initializing the detection space according to a preset mean strategy to obtain a first combination instance.

[0069] It should be noted that in this step, in order to provide a priori estimation, the system will use a preset mean strategy to initialize the detection space and generate a first combination instance. These instances include anchors that are evenly distributed in the detection space and have a preset size, shape, and orientation. Specifically, the system determines the number of anchors that need to be generated based on the size and complexity of the detection space. Use a clustering algorithm such as k-means to divide the detection space so that the anchors can be evenly distributed in the detection space.

[0070] Step A50: adding noise according to the real labeled data to obtain a second combination instance.

[0071] It should be noted that if Figure 2 As shown, in this step, the system randomly selects a part of the data from the real annotated data set as the basic sample, and adds random noise to the selected basic sample. Noise can include position noise, size noise, speed noise, etc., and the specific form can be set according to actual needs. For example, for a vehicle target, we can add a certain position noise (such as ±0.5 meters) based on its true position, add ±5% noise to the size, and add ±1m / s noise to the speed. In addition, it should be noted that adding noise is a commonly used data enhancement method, which can enable the model to learn more generalized feature representations during training. It can be understood that by adding noise to the real annotated data to generate a second combination instance, the robustness and generalization ability of the model are increased.

[0072] Step S20: obtaining a target combination instance by processing according to the camera parameters, the multi-view image information, the historical combination instance, the first combination instance, and the second combination instance.

[0073] It should be noted that in this step, Figure 2 As shown in the figure, the system generates the final target combination instance by fusing information from multiple sources. Specifically, the system uses the image backbone network to extract features from multi-view image information, and supervises through the perspective 3D detection head to improve detection accuracy. Then, the camera pose embedding module is used to fuse the camera parameters to generate a unified feature representation. Finally, the historical combination instance, the first combination instance, and the second combination instance are combined, and the temporal features are fused through the current cycle layer and the cross-cycle layer to generate the final target combination instance.

[0074] It is understandable that time series feature fusion can improve detection accuracy and stability. By combining historical information with current information, the model can better understand the dynamic changes of the target and improve the accuracy and robustness of detection.

[0075] Step S30: Decode the target combination instance to achieve target detection.

[0076] It should be noted that the decoder is a network structure that maps a high-dimensional feature vector to a specific target detection result. In this embodiment, the decoder is responsible for decoding the "anchor+feature" combination instance output by the cross-cycle layer into a target detection result.

[0077] In addition, it should be noted that in this step, Anchors and Instance Features are processed and used for classification and regression tasks. Figure 3 As shown in the figure, ReLU / GELU is a nonlinear activation function layer, and Linear is a linear layer. The top-k strategy is used to retain or delete all the detection results, and the retained instances enter the subsequent algorithm pipeline as historical time series instances. Finally, the category, confidence, and Boxes (3D detection box of the target instance) are output.

[0078] For Box decoding, after adopting the top-k strategy, the output is x, y, z, w, l, h, sin yaw ,cos yaw The multi-dimensional tensor needs to be decoded to get the final actual result. Specifically, the final output Boxes position coordinates are X, Y, Z, as shown in Formula 1.

[0079] [X, Y, Z] = [x, y, z] (Formula 1)

[0080] The values ​​of the output dimensions W, L, and H need to be converted to actual dimensions through an exponential function, because during training or processing, these dimensions are converted to logarithmic space for numerical stability, as shown in Formula 2.

[0081] [W,L,H]=[e w ,e l ,e h ](Formula 2)

[0082] To obtain the final yaw value of the output, it is necessary to calculate the inverse tangent value of the two components sin(yaw) and cos(yaw) to obtain the angle yaw, as shown in Formula 3.

[0083] |yaw|=|atan2(sin yaw ,cos yaw )|(Formula 3)

[0084] It is understandable that this design completely decouples the structured state of image features and instances, making it easier to add prior knowledge and extend to time series tracking. Here, the function returns an angle value in radians with a range of (-π,π], which can correctly locate the angle to any of the four quadrants based on the signs of sin(yaw) and cos(yaw).

[0085] The present embodiment provides a method for target detection with perspective supervision, and the method of the present embodiment includes: obtaining camera parameters, multi-view image information, historical combination instances, first combination instances, and second combination instances, wherein the combination instance includes anchor points and their corresponding features; processing the target combination instance according to the camera parameters, the multi-view image information, the historical combination instance, the first combination instance, and the second combination instance; decoding the target combination instance to achieve target detection. In summary, the present embodiment improves the adaptability of the image backbone to 3D target detection tasks through perspective view supervision, and adopts an anchor-based sparse BEV detector scheme, which is more flexible in temporal fusion, solves the problem of how to improve the flexibility and accuracy of target detection, and improves detection performance and reasoning speed.

[0086] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the above-mentioned embodiment 1 can refer to the above introduction, and will not be repeated later. Figure 4 , Figure 4 This is a flow chart of the second embodiment of the target detection method with perspective supervision of the present application, wherein step S20 specifically includes:

[0087] Step S201: obtaining multi-view image features, detection results, and perspective 3D target anchor points according to the camera parameters, the multi-view image information, and the first combination instance.

[0088] It should be noted that multi-view image features refer to feature maps obtained after feature extraction of images captured from different camera perspectives (i.e., multi-view image information), which contain high-order semantic information of the image; perspective 3D target anchors refer to the initial position and posture information of the 3D target box obtained by perspective detection, which is used for subsequent target detection and tracking; the detection result refers to the "anchor + feature" obtained by fusing various types of data in this cycle layer.

[0089] In a feasible implementation manner, the step S201 specifically includes:

[0090] Step C10: extracting image features from the multi-view images to obtain multi-view image features.

[0091] It should be noted that multi-view image features refer to the deep features extracted from images captured from different camera perspectives, which together describe the three-dimensional information of the vehicle's surroundings; image feature extraction is to process multi-view images captured by multiple cameras using a deep learning image backbone network (such as ResNet, VGG, etc.). Specifically, Figure 2 As shown in the figure, the system takes images captured by multiple cameras at different positions of the vehicle as input and preprocesses the input images, including resizing and normalization, to meet the input requirements of the image backbone network. The preprocessed images are sent to the pre-trained image backbone network, and the deep features of the images (these features contain high-level semantic information in the image) are extracted layer by layer through the convolutional layer, pooling layer and other structures in the network. Finally, the image features from different camera perspectives are integrated to form multi-perspective image features.

[0092] Step C20: Supervise the image feature extraction process to obtain the perspective 3D target anchor point.

[0093] It should be noted that the perspective 3D target anchor refers to the prior box generated based on the supervisory signal for subsequent BEV target detection under the perspective view. Figure 2 As shown, in this step, the system connects a perspective 3D detection head after the image backbone network, and the detection head is used to detect 3D targets in the image under a perspective view.

[0094] Step C30: Obtaining a detection result according to the multi-view image features, the camera parameters and the first combination instance.

[0095] It should be noted that in this step, the system will first map the camera parameters into a high-dimensional feature vector, and perform time-series fusion of it with the high-dimensional features of the target (i.e., the features of the first combination instance) and the multi-view image features in this cycle to obtain a set of "anchor + feature" detection results.

[0096] In a feasible implementation manner, the step C20 specifically includes:

[0097] Step C201: Supervise the image feature extraction process according to a first preset supervision method to obtain an initial perspective 3D target anchor point.

[0098] It should be noted that in this step, the first preset supervision method is to supervise the image feature extraction process through a perspective 3D detection head (auxiliary detection head). Specifically, the system receives image features as input according to the perspective 3D detection head, and outputs 3D target detection results under the perspective view, including the initial position (2D bbox), size, orientation and other information of the target. This information constitutes the initial perspective 3D target anchor point, which provides a basic target position and posture estimation for subsequent steps. It can be understood that this step performs additional supervision on image features through the perspective 3D detection head, so that the image feature extraction process focuses more on the detection task of 3D targets, thereby improving the accuracy of target detection.

[0099] Step C202: Supervise the image feature extraction process according to a second preset supervision method to obtain depth information.

[0100] It should be noted that in this step, the second preset supervision method is to use dense depth supervision of LiDAR point cloud. Specifically, the system will use the precise depth information provided by LiDAR point cloud to perform additional depth supervision on the image feature extraction process. This can be achieved by projecting the LiDAR point cloud onto the image plane to form a depth map and fusing or comparing it with the image features. Through dense depth supervision of LiDAR point cloud, the accuracy of image features in the depth dimension can be significantly improved, thereby improving the accuracy of perspective 3D target detection.

[0101] Step C203: obtaining a perspective 3D target anchor point according to the initial perspective 3D target anchor point and the depth information.

[0102] It should be noted that in this step, the system will fuse the initial perspective 3D target anchor point obtained in step C201 with the depth information obtained in step C202 to obtain a more accurate perspective 3D target anchor point. Specifically, the position (2D bbox), size, orientation and other information of the initial perspective 3D target anchor point are combined with the depth information to form a perspective 3D target anchor point containing complete 3D position and posture information. It can be understood that the perspective 3D target anchor point not only contains the position information of the target on the image plane, but also contains its depth information in three-dimensional space.

[0103] In a feasible implementation manner, the step C30 specifically includes:

[0104] Step C301: mapping the camera parameters according to a preset position conversion strategy to obtain a high-dimensional feature vector.

[0105] It should be noted that if Figure 2As shown in the figure, the preset position conversion strategy refers to the pre-set Camera Pose Embedding module, which uses Fourier encoding technology to map camera parameters (including the position and orientation of the camera) into high-dimensional feature vectors, and can convert camera parameters from physical space to feature space. Specifically, the system receives input camera parameters, including the position and orientation of the camera. Then, Fourier transform is used to map the camera parameters to the high-dimensional feature space to generate the corresponding high-dimensional feature vector.

[0106] Step C302: performing feature combination according to the high-dimensional feature vector and the first combination instance to obtain a third combination instance.

[0107] It should be noted that in this step, the system will perform operations such as concatenation or weighted summation of the high-dimensional feature vector and the features in the first combination instance (i.e., the initialized anchor point and features) to obtain a new feature combination, i.e., the third combination instance. The third combination instance is used as input for subsequent processing to generate the final detection result.

[0108] Step C303: performing feature sampling and weighted summation on the camera parameters, the third combination instance, and the multi-view image features according to a preset sampling strategy to obtain a detection result.

[0109] It should be noted that, specifically, in this step, the system will sample and weighted sum the multi-view image features according to a preset sampling strategy (such as the grid_sample sampling algorithm or the LSS algorithm) to obtain fused image features. The fused image features are then fused again with the third combination instance, while considering the influence of camera parameters (such as by feature combination or weighting), to obtain the final detection result. In addition, it should be noted that the preset sampling strategy refers to a rule or method for extracting useful information from input data. In this embodiment, a sampling strategy based on the grid_sample sampling algorithm or the LSS algorithm is used to extract useful information from multi-view image features.

[0110] Step S202: Processing is performed according to the detection result, the historical combination instance, the second combination instance, the multi-view image feature and the perspective 3D target anchor point to obtain a target combination instance.

[0111] It should be noted that, specifically, Figure 2As shown in the figure, the system will input the detection results of this cycle, historical combination instances, second combination instances, multi-view image features, and perspective 3D target anchors into the cross-cycle layer through cross-cycle. The cross-cycle layer effectively fuses information from different sources through mechanisms such as cross attention, self-attention, and feature aggregation to generate more accurate and robust target combination instances. It can be understood that cross-cycle fusion refers to the process of effectively integrating information from different time points, which can improve the accuracy and stability of target detection.

[0112] In this embodiment, by designing the camera extrinsic Fourier encoding module, the generalization ability of the camera extrinsic parameters is improved, and the accuracy and robustness of target detection are further improved. At the same time, by performing additional supervision through the perspective 3D detection head and dense depth supervision of the LiDAR point cloud, the accuracy of the image features in the depth dimension is improved, thereby improving the accuracy of 3D target detection.

[0113] This application also provides a target detection device with perspective supervision, please refer to Figure 5 , the target detection device with perspective supervision comprises:

[0114] A data acquisition module 10 is used to acquire camera parameters, multi-view image information, a historical combination instance, a first combination instance, and a second combination instance, wherein the combination instance includes an anchor point and its corresponding features;

[0115] A data processing module 20, configured to obtain a target combination instance by processing according to the camera parameters, the multi-view image information, the historical combination instance, the first combination instance, and the second combination instance;

[0116] The target detection module 30 is used to decode the target combination instance to achieve target detection.

[0117] The target detection device with perspective supervision provided by the present application adopts the target detection method with perspective supervision in the above-mentioned embodiment, which can solve the technical problem of how to improve the flexibility and accuracy of target detection. Compared with the prior art, the beneficial effects of the target detection device with perspective supervision provided by the present application are the same as the beneficial effects of the target detection method with perspective supervision provided by the above-mentioned embodiment, and the other technical features in the target detection device with perspective supervision are the same as the features disclosed in the above-mentioned embodiment method, which will not be repeated here.

[0118] In one embodiment, the data acquisition module 10 is further used to acquire the motion posture of the vehicle and an initial historical combination instance; and to perform coordinate transformation on the initial historical combination instance according to the motion posture of the vehicle to obtain a historical combination instance.

[0119] In one embodiment, the data acquisition module 10 is further used to acquire a detection space and real labeled data; initialize the detection space according to a preset mean strategy to obtain a first combination instance; and add noise according to the real labeled data to obtain a second combination instance.

[0120] In one embodiment, the data processing module 20 is further used to obtain multi-view image features, detection results and perspective 3D target anchor points based on the camera parameters, the multi-view image information and the first combination instance; and to obtain a target combination instance by processing the detection results, the historical combination instance, the second combination instance, the multi-view image features and the perspective 3D target anchor points.

[0121] In one embodiment, the data processing module 20 is also used to extract image features from the multi-view image to obtain multi-view image features; supervise the image feature extraction process to obtain perspective 3D target anchor points; and obtain detection results based on the multi-view image features, the camera parameters and the first combination instance.

[0122] In one embodiment, the data processing module 20 is also used to supervise the image feature extraction process according to a first preset supervision method to obtain an initial perspective 3D target anchor point; supervise the image feature extraction process according to a second preset supervision method to obtain depth information; and obtain a perspective 3D target anchor point based on the initial perspective 3D target anchor point and the depth information.

[0123] In one embodiment, the data processing module 20 is further used to map the camera parameters according to a preset position conversion strategy to obtain a high-dimensional feature vector; perform feature combination based on the high-dimensional feature vector and the first combination instance to obtain a third combination instance; perform feature sampling and weighted summation on the camera parameters, the third combination instance, and the multi-view image features according to a preset sampling strategy to obtain a detection result.

[0124] The present application provides a target detection device with perspective supervision, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the target detection method with perspective supervision in the above-mentioned embodiment one.

[0125] Reference below Figure 6, which shows a schematic diagram of the structure of a target detection device with perspective supervision suitable for implementing the embodiment of the present application. The target detection device with perspective supervision in the embodiment of the present application may include but is not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 6 The target detection device with perspective supervision shown is merely an example and should not bring any limitation to the functionality and scope of use of the embodiments of the present application.

[0126] like Figure 6 As shown, the target detection device with perspective supervision may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the target detection device with perspective supervision are also stored. The processing device 1001, ROM1002, and RAM1004 are connected to each other via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices 1003 including, for example, a magnetic tape, a hard disk, etc.; and communication devices 1009. The communication device 1009 can allow the target detection device with perspective supervision to communicate with other devices wirelessly or by wire to exchange data. Although the target detection device with perspective supervision having various systems is shown in the figure, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems may be implemented or have instead.

[0127] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0128] The target detection device with perspective supervision provided by the present application adopts the target detection method with perspective supervision in the above embodiment, which can solve the technical problem of how to improve the flexibility and accuracy of target detection. Compared with the prior art, the beneficial effects of the target detection device with perspective supervision provided by the present application are the same as the beneficial effects of the target detection method with perspective supervision provided by the above embodiment, and the other technical features in the target detection device with perspective supervision are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0129] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0130] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0131] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer programs) stored thereon, the computer-readable program instructions being used to execute the target detection method with perspective supervision in the above-mentioned embodiment.

[0132] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.

[0133] The computer-readable storage medium may be included in the target detection device with perspective supervision; or may exist independently without being assembled into the target detection device with perspective supervision.

[0134] The above-mentioned computer-readable storage medium carries one or more programs. When the above-mentioned one or more programs are executed by a target detection device with perspective supervision, the target detection device with perspective supervision: obtains camera parameters, multi-view image information, historical combination instances, first combination instances and second combination instances, and the combination instance includes an anchor point and its corresponding features; processes according to the camera parameters, the multi-view image information, the historical combination instance, the first combination instance, and the second combination instance to obtain a target combination instance; and decodes the target combination instance to achieve target detection.

[0135] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0136] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0137] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.

[0138] The readable storage medium provided by the present application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned target detection method with perspective supervision, and can solve the technical problem of how to improve the flexibility and accuracy of target detection. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided by the present application are the same as the beneficial effects of the target detection method with perspective supervision provided by the above-mentioned embodiment, and will not be repeated here.

[0139] The present application also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the target detection method with perspective supervision as described above.

[0140] The computer program product provided by this application can solve the technical problem of how to improve the flexibility and accuracy of target detection. Compared with the prior art, the beneficial effects of the computer program product provided by this application are the same as the beneficial effects of the target detection method with perspective supervision provided by the above embodiment, which will not be repeated here.

[0141] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A method for object detection with perspective supervision, characterized in that: The method comprises: Obtaining camera parameters, multi-view image information, a historical combination instance, a first combination instance, and a second combination instance, wherein the combination instance includes an anchor point and features corresponding thereto; Processing is performed according to the camera parameters, the multi-view image information, the historical combination instance, the first combination instance, and the second combination instance to obtain a target combination instance; The target combination instance is decoded to achieve target detection.

2. The method according to claim 1, characterized in that The step of obtaining a target combination instance by processing according to the camera parameters, the multi-view image information, the historical combination instance, the first combination instance, and the second combination instance comprises: Obtain multi-view image features, detection results, and perspective 3D target anchor points according to the camera parameters, the multi-view image information, and the first combination instance; A target combination instance is obtained by processing according to the detection result, the historical combination instance, the second combination instance, the multi-view image feature and the perspective 3D target anchor point.

3. The method according to claim 2, characterized in that The step of obtaining multi-view image features, detection results, and perspective 3D target anchor points according to the camera parameters, the multi-view image information, and the first combination instance comprises: Extracting image features from the multi-view images to obtain multi-view image features; Supervise the image feature extraction process to obtain perspective 3D target anchor points; A detection result is obtained according to the multi-view image features, the camera parameters and the first combination instance.

4. The method according to claim 3, characterized in that The step of supervising the image feature extraction process to obtain the perspective 3D target anchor point comprises: Supervising the image feature extraction process according to a first preset supervision method to obtain an initial perspective 3D target anchor point; Supervising the image feature extraction process according to a second preset supervision method to obtain depth information; A perspective 3D object anchor point is obtained according to the initial perspective 3D object anchor point and the depth information.

5. The method according to claim 3, characterized in that The step of obtaining the detection result according to the multi-view image features, the camera parameters and the first combination example comprises: Mapping the camera parameters according to a preset position conversion strategy to obtain a high-dimensional feature vector; Performing feature combination according to the high-dimensional feature vector and the first combination instance to obtain a third combination instance; According to a preset sampling strategy, feature sampling and weighted summation are performed on the camera parameters, the third combination instance, and the multi-view image features to obtain a detection result.

6. The method according to claim 1, characterized in that The steps to obtain a historical combination instance include: Obtain the vehicle's motion posture and initial history combination instance; The initial historical combination instance is coordinate-converted according to the vehicle motion posture to obtain a historical combination instance.

7. The method according to any one of claims 1 to 6, characterized in that The steps of obtaining the first combination instance and the second combination instance include: Obtain detection space and real annotation data; Initializing the detection space according to a preset mean strategy to obtain a first combination instance; Noise is added according to the real labeled data to obtain a second combination instance.

8. An object detection device with perspective supervision, characterized in that The device comprises: A data acquisition module, used to acquire camera parameters, multi-view image information, historical combination instances, first combination instances, and second combination instances, where the combination instances include anchor points and their corresponding features; A data processing module, configured to obtain a target combination instance by processing according to the camera parameters, the multi-view image information, the historical combination instance, the first combination instance, and the second combination instance; The target detection module is used to decode the target combination instance to achieve target detection.

9. An object detection device with perspective supervision, characterized in that The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the target detection method with perspective supervision as described in any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the target detection method with perspective supervision as described in any one of claims 1 to 7 are implemented.

Images