A target identification method, device, equipment, storage medium and product

Abstract

The application discloses a target identification method and device, equipment, storage medium and product, relates to the technical field of image recognition, and according to the set target screening rule, a target candidate set is screened from a feature map set to exclude irrelevant targets. The feature information of each candidate target in the target candidate set is semantically aligned with the ego vehicle intention vector to obtain a fusion feature vector. The visual representation features of each candidate target and the ego vehicle intention vector are analyzed to obtain an intention matching score of each candidate target. Based on the intention matching score of each candidate target and the rule score, pseudo-label data is constructed. According to the pseudo-label data and the identification result output by analyzing the fusion feature vector by using the fusion classifier, the parameters of the fusion classifier are adjusted to obtain a trained fusion classifier. Through weak supervision training, the generalization ability of the fusion classifier in a complex environment is improved. The trained fusion classifier can accurately identify the set target.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, and in particular to a target recognition method and device, equipment, storage medium and product. BACKGROUND

[0002] As a key node with high traffic density and complex traffic rules, the perception and decision performance of the urban intersection is directly related to traffic safety and system traffic efficiency. In typical urban roads, small-sized targets such as traffic lights, signboards, and road cones are widely distributed in important locations such as intersections, construction areas, and ramp boundaries. These targets, although occupying only a very small pixel area in the image, carry vital road traffic information.

[0003] Traditional target recognition methods are difficult to adapt flexibly in dynamic traffic scenes, especially in multi-light overlapping, semantically ambiguous, or long-distance scenes, the target traffic light recognition accuracy is low, and the perception generalization ability is poor. And the traditional target recognition method lacks effective fusion of traffic light behavior guidance. Especially in the complex urban intersection scene, there are many traffic lights, and there are problems such as serious occlusion, small size, and long distance, and the system often has difficulty in accurately determining which traffic light the vehicle should follow.

[0004] Therefore, how to improve the accuracy of traffic light recognition is a problem to be solved by those skilled in the art. SUMMARY

[0005] The present application provides a target recognition method, device, equipment, storage medium and product to at least solve the problem that it is difficult to accurately determine which traffic light the vehicle should follow in the related art.

[0006] The present application provides a target recognition method, comprising:

[0007] According to the set target screening rule, a target candidate set is screened from the feature map set; wherein the target screening rule is set according to the relevance between the set target and the ego vehicle driving behavior; the target candidate set contains each candidate target and its corresponding rule score;

[0008] The feature information of each candidate target in the target candidate set is semantically aligned with the ego vehicle intention vector to obtain a fusion feature vector;

[0009] The visual representation features of each candidate target and the ego vehicle intention vector are analyzed to obtain an intention matching score of each candidate target;

[0010] Based on the intention matching score of each candidate target and the rule score, pseudo-label data is constructed; wherein the pseudo-label data contains the predicted probability of each candidate target belonging to the set target;

[0011] The parameters of the fusion classifier are adjusted according to the pseudo label data and the recognition result output by the fusion classifier analyzing the fusion feature vector, so as to obtain a trained fusion classifier, so as to identify the set target by using the trained fusion classifier.

[0012] The application further provides a target identification device, comprising a screening unit, an alignment unit, a matching unit, a construction unit and an adjustment unit.

[0013] The screening unit is configured to screen a target candidate set from the feature map set according to a set target screening rule; wherein the target screening rule is set according to the relevance between the set target and the driving behavior of the ego vehicle; and the target candidate set comprises each candidate target and a corresponding rule score thereof.

[0014] The alignment unit is configured to perform semantic alignment between the feature information of each candidate target in the target candidate set and the ego vehicle intention vector, so as to obtain a fusion feature vector.

[0015] The matching unit is configured to analyze the visual representation feature of each candidate target and the ego vehicle intention vector, so as to obtain an intention matching score of each candidate target.

[0016] The construction unit is configured to construct pseudo label data based on the intention matching score of each candidate target and the rule score; wherein the pseudo label data comprises a predicted probability that each candidate target belongs to the set target.

[0017] The adjustment unit is configured to adjust the parameters of the fusion classifier according to the pseudo label data and the recognition result output by the fusion classifier analyzing the fusion feature vector, so as to obtain a trained fusion classifier, so as to identify the set target by using the trained fusion classifier.

[0018] The application further provides an electronic device, comprising a memory for storing a computer program, and a processor for executing the computer program to implement the steps of any of the above target identification methods.

[0019] The application further provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of any of the above target identification methods.

[0020] The application further provides a computer program product, comprising a computer program, and the computer program is executed by a processor to implement the steps of any of the above target identification methods.

[0021] This application uses a set of target candidates to filter from a feature map set according to a set of defined target selection rules. These rules are based on the correlation between the defined target and the vehicle's driving behavior. The target candidate set includes each candidate target and its corresponding rule score. By setting the target selection rules, irrelevant targets in the feature map set can be eliminated. To increase the accuracy of target selection, the feature information of each candidate target in the target candidate set can be semantically aligned with the vehicle's intent vector to obtain a fused feature vector. The visual representation features of each candidate target and the vehicle's intent vector are analyzed to obtain the intent matching score for each candidate target. Based on the intent matching score and rule score of each candidate target, pseudo-label data is constructed, which includes the predicted probability that each candidate target belongs to the defined target. Based on the pseudo-label data and the recognition results output by the fusion classifier in analyzing the fused feature vector, the parameters of the fusion classifier are adjusted to obtain a trained fusion classifier, which can then be used to identify the defined target. In this application, based on the selection of a candidate target set, a semantic alignment mechanism between the vehicle's intent and the candidate targets is established through intent guidance to address the problem of discriminating between multiple candidate targets in the same scene that are related to the vehicle's behavior but have different priorities. A pseudo-label-based weakly supervised training mechanism fuses the rule scores and intent matching scores of candidate targets to generate a pseudo-label distribution. By weakly supervising the fusion classifier, the difficulties in constructing datasets and the scarcity of true labels in object detection and recognition tasks are solved. Furthermore, the pseudo-label-based weakly supervised training mechanism improves the generalization ability of the fusion classifier in complex environments. The trained fusion classifier can accurately identify the designated targets. Attached Figure Description

[0022] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a target recognition method provided for an embodiment of this application;

[0024] Figure 2 A flowchart illustrating a method for selecting a target candidate set from a feature map set, provided in an embodiment of this application;

[0025] Figure 3 This application provides a schematic diagram of a target traffic light selection process.

[0026] Figure 4 This is a schematic diagram of the structure of a target recognition device provided in an embodiment of this application. Detailed Implementation

[0027] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0028] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., in this application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0029] While significant progress has been made in traffic light recognition and behavioral decision guidance, current solutions still have significant shortcomings in real-world urban road environments. Firstly, in traffic light recognition, complex urban intersections often present multiple traffic light targets, which are small, distant, and obstructed. Current methods struggle to accurately identify the target traffic light by considering the vehicle's intent—the traffic light the vehicle should follow. Furthermore, attention is frequently diverted to multiple irrelevant light states during traffic light recognition, leading to misjudgments and delayed responses. Secondly, regarding the coordination between perception and decision-making modules, most system modules are fragmented, lacking effective information sharing mechanisms. This results in a lack of intent and behavioral guidance for the detection and recognition of the primary target.

[0030] To enhance the spatial understanding capabilities of perception systems, bird's-eye view (BEV) has become the mainstream form of perception modeling. BEV constructs a scale-consistent, geometrically aligned spatial semantic representation by projecting the views of multiple cameras onto a unified ground plane. Despite BEV's advantages in spatial semantic representation, an effective and systematic solution remains lacking for efficiently and reliably extracting key traffic light states from BEV feature maps and driving behavior prediction.

[0031] Therefore, the embodiment of the present application provides a target identification method, device, equipment, storage medium and product. The target screening rule is set by setting the relevance of the target and the driving behavior of the ego vehicle, and the target candidate set is screened from the feature map set according to the set target screening rule, so as to exclude irrelevant targets in the feature map set. On this basis, the semantic alignment mechanism between the ego vehicle intention and the candidate target is established through the intention guidance, so as to solve the discrimination problem that multiple candidate targets in the same scene are related to the ego vehicle behavior but have different priorities. Based on the weak supervision training mechanism of the pseudo label, the rule score and the intention matching score of the candidate target are fused to generate a pseudo label distribution. Through the weak supervision training of the fusion classifier, the problem of difficult construction of the data set and scarcity of real labels in the target detection and identification task is solved. Moreover, through the weak supervision training mechanism based on the pseudo label, the generalization ability of the fusion classifier in the complex environment is improved, so that the trained fusion classifier can accurately identify the set target.

[0032] The small target mentioned in the embodiment of the present application refers to a target that presents a scale significantly smaller than the average target size in the image space or the BEV space, and the semantic structure in the feature map is easy to lose and the detection difficulty is high. The small target usually includes but is not limited to: traffic lights, road cones, road signs, warning signs, temporary obstacles and other small volume, weak texture and frequently occluded structural units. In the subsequent content, traffic lights are taken as examples for expansion.

[0033] In order for those skilled in the art to better understand the present application, the present application will be further described in detail below in combination with the drawings and specific embodiments.

[0034] Figure 1 A flowchart of a target identification method provided by the embodiment of the present application includes:

[0035] S101: According to the set target screening rule, a target candidate set is screened from the feature map set.

[0036] The target screening rule can be set according to the relevance of the set target and the driving behavior of the ego vehicle.

[0037] In actual application, the feature map set can be extracted from the bird's eye view feature map. The bird's eye view feature map, i.e. the BEV feature map, is a feature map obtained by bird's eye view coding of multiple camera images.

[0038] The target candidate set includes each candidate target and the corresponding rule score of the candidate target.

[0039] In the embodiment of the present application, traffic lights required to be recognized by a vehicle are taken as examples for expansion, and the set target is a target traffic light. The target traffic light refers to a traffic light that should be followed by the ego vehicle.

[0040] In order to effectively reduce the search space, eliminate irrelevant lights, and improve the identification efficiency of the target traffic light, a target screening rule can be set according to the relevance of the traffic light and the driving behavior of the ego vehicle.

[0041] The target screening rule can include type filtering and spatial relationship judgment. By comprehensively considering the type of traffic light and the spatial relationship between the traffic light and the ego vehicle, the irrelevant interference targets unrelated to the behavior of the ego vehicle can be quickly excluded. A feasible screening method can be referred to the introduction of Figure 2 , which will not be described here.

[0042] S102: Align the feature information of each candidate target in the target candidate set with the ego vehicle intention vector to obtain a fusion feature vector.

[0043] On the basis of screening the target candidate set, a semantic alignment mechanism between the ego vehicle intention and the candidate traffic light is established through the intention guiding module to solve the discrimination problem that multiple candidate lights are related to the behavior of the ego vehicle but have different priorities in the same scene. The priority refers to the degree of influence of multiple candidate lights on the behavior of the ego vehicle, and the traffic light required to be followed in the driving process of the ego vehicle has the highest priority.

[0044] In order to obtain a more expressive intention vector of the ego vehicle, the trajectory point sequence of the ego vehicle is used as input, and the encoder in the deep learning model (Transformer) based on multi-head self-attention mechanism is used to obtain the intention vector of the ego vehicle. By splicing the feature information of each candidate target with the intention vector of the ego vehicle, a fusion feature vector is obtained.

[0045] In specific implementation, the trajectory point sequence of the ego vehicle can be extracted through a path planning system or a multi-modal trajectory prediction module: .

[0046] In the trajectory encoder, in order to obtain a more expressive intention vector, the trajectory point sequence G of the ego vehicle is used as input, and the Transformer encoder is used to obtain the ego vehicle intention embedding vector. The operation performed by the Transformer encoder can be denoted as IntentEncoder: ; wherein F intent represents the intention vector of the ego vehicle.

[0047] The construction process of IntentEncoder is as follows: first, the trajectory point sequence G is mapped into an embedding vector sequence ; wherein, , represents the spatial position of the candidate target, and LinearEmbedding represents linear embedding processing.

[0048] Position encoding is added to inject position information, so that the model can perceive the time sequence of points: ; wherein represents a position code.

[0049] The input is fed into a multi-layer Transformer Encoder for self-attention modeling to capture the dependencies within the sequence. Finally, a fixed-dimensional intent vector is generated by a pooling operation or by taking the last token .

[0050] Extracting feature information of each candidate target from the target candidate set can include visual representation features , direction angle cosine values , and relative distances of candidate targets and the ego vehicle ; wherein P ego represents the spatial position of the ego vehicle.

[0051] The ego vehicle intent vector is fused with the feature information of each candidate target to obtain an intent vector of each candidate target.

[0052] ; wherein z i represents a fused vector obtained by concatenation, which fuses the candidate light information and the intent vector.

[0053] Then the intent vector of the candidate light is generated: ;

[0054] wherein is a linear mapping weight matrix, projecting the fused vector z i to a query space, is the intent vector of the i-th candidate light.

[0055] According to the key-value pair information of each position in the bird's eye view feature map, the intent vector of each candidate target is aligned with the ego vehicle intent vector to obtain a fused feature vector.

[0056] To strengthen the alignment process of the candidate light and the ego vehicle intent, a cross-attention mechanism is used to guide the perception layer to focus on the region related to the current intent, and the fused feature vector is calculated as: ; ;

[0057] wherein F i represents the fused feature vector, represents the intent vector of the i-th candidate light, represents the key (key) of each position in the corresponding BEV feature map, represents the value (value) of each position in the corresponding BEV feature map.

[0058] Cross-Attention can adopt a multi-head mechanism, allowing different intention subspaces to participate in attention calculation respectively. i reflects the correlation between the candidate light and the ego vehicle intention vector.

[0059] S103: Analyze the visual representation features of each candidate target and the ego vehicle intention vector to obtain an intention matching score of each candidate target.

[0060] In a specific implementation, the visual representation features of each candidate target and the ego vehicle intention vector can be spliced to obtain a spliced vector; the spliced vector is analyzed by using a multi-layer fully connected layer, and the analysis result is mapped to an intention matching score.

[0061] For each traffic light L i , the intention matching score is obtained by the visual representation features in the BEV feature map thereof and the intention vector of the ego vehicle : ;

[0062] wherein, represents a splicing operation or other feature fusion manner, and the Sigmoid function maps the output to interval, representing the matching probability.

[0063] S104: Construct pseudo-label data based on the intention matching score of each candidate target and the rule score.

[0064] wherein, the pseudo-label data contains the predicted probability of each candidate target belonging to the set target.

[0065] In the embodiments of the present application, the intention matching score of each candidate target and the rule score can be subjected to probability normalization processing to obtain the predicted probability of each candidate target belonging to the set target. That is, the pseudo-label data is the normalized probability distribution.

[0066] In order to further improve the generalization of the model, the present application proposes a pseudo-label generation strategy that fuses rule scoring and intention matching score , and the pseudo-label is a normalized probability distribution: ;

[0067] wherein, represents the predicted probability of candidate target i belonging to the set target, is a weight adjustment hyperparameter, which can be set to , and N is the number of targets.

[0068] S105: Adjusting parameters of the fusion classifier according to the pseudo label data and the recognition result output by analyzing the fusion feature vector by the fusion classifier, to obtain a trained fusion classifier, so as to identify the set target by using the trained fusion classifier.

[0069] After obtaining the fusion feature vector, the fusion feature vector can be input into the fusion classifier to determine the probability of each candidate target belonging to the set target.

[0070] Taking the candidate target as a candidate traffic light as an example, the fusion classifier is used to jointly infer all candidate traffic lights, and the confidence of each candidate traffic light being a "target traffic light" is output. First, all fusion feature vectors, i.e., intention guidance perception outputs F i are input into the fusion classifier, and the probability of each candidate traffic light being a target traffic light is output

[0071] Classifier is one or more fully connected layers, and Softmax converts the classification result into a normalized probability.

[0072] The fusion classifier supports information interaction and semantic comparison between different candidate traffic lights, and improves the discrimination robustness under complex intersection conditions such as mixed multiple traffic lights and occlusion interference.

[0073] In a specific implementation, a cross-entropy loss function can be constructed according to the probability of each candidate target belonging to the set target and the predicted probability of each candidate target belonging to the set target; and the parameters of the fusion classifier are adjusted by using the cross-entropy loss function to obtain a trained fusion classifier.

[0074] The fusion classifier is trained by using the following cross-entropy loss function:

[0075]

[0076] wherein L cls represents the cross-entropy loss function.

[0077] In the embodiments of the present application, based on the weak supervision training mechanism of the pseudo label, the rule score and the intention matching score of the candidate light are weighted and fused to generate a pseudo label distribution. Through weak supervision training of the fusion classifier, the problem of difficult construction of a traffic light data set and scarcity of real labels in the target detection and recognition task is solved. Moreover, through the weak supervision training mechanism based on the pseudo label, the generalization ability of the model in unstructured intersections / complex environments is improved. By limiting the non-differentiable rule in the candidate light set screening and pseudo label generation stage, and training the differentiable model in a weak supervision manner, the problem of direct interference of the non-differentiable rule with gradient propagation is effectively avoided, which is helpful for the system to perform end-to-end training in a planning-oriented manner.​​​

[0078] According to the technical solution, the target candidate set is filtered from the feature map set according to the set target filtering rule. The target filtering rule is set according to the relevance between the set target and the driving behavior of the ego vehicle. The target candidate set includes each candidate target and the corresponding rule score. By setting the target filtering rule, irrelevant targets in the feature map set can be excluded. To increase the correctness of target selection, the feature information of each candidate target in the target candidate set can be semantically aligned with the ego vehicle intention vector to obtain a fusion feature vector. The visual representation feature of each candidate target and the ego vehicle intention vector are analyzed to obtain an intention matching score of each candidate target. Based on the intention matching score and the rule score of each candidate target, pseudo-label data is constructed. The pseudo-label data includes the predicted probability of each candidate target belonging to the set target. According to the pseudo-label data and the recognition result output by analyzing the fusion feature vector by the fusion classifier, the parameters of the fusion classifier are adjusted to obtain a trained fusion classifier, so as to identify the set target by using the trained fusion classifier. In this application, based on the filtered target candidate set, a semantic alignment mechanism between the ego vehicle intention and the candidate target is established by intention guidance to solve the problem of distinguishing multiple candidate targets with different priorities in the same scene. Based on the weak supervision training mechanism of the pseudo-label, the rule score and the intention matching score of the candidate target are fused to generate a pseudo-label distribution. By weakly supervising the fusion classifier, the problem of difficult data set construction and lack of real labels in the target detection and recognition task is solved. Moreover, the weak supervision training mechanism based on the pseudo-label improves the generalization ability of the fusion classifier in complex environments. The trained fusion classifier can accurately identify the set target.

[0079] Figure 2 A flowchart of a method for filtering a target candidate set from a feature map set is provided in the embodiments of the present application. The method includes:

[0080] S201: Extract a feature map set from the bird's eye view feature map by using a parallel branch network.

[0081] The bird's eye view feature map is a feature map obtained by encoding a bird's eye view map from multiple camera images.

[0082] The feature map set can include a feature map set corresponding to a first target and a feature map set corresponding to a second target. The number of pixel points of the first target is greater than the number of pixel points of the second target.

[0083] In a specific implementation, the first branch network can be used to perform feature extraction on the bird's eye view feature map to obtain a first target feature map set corresponding to each first target; the second branch network can be used to perform feature extraction on the bird's eye view feature map to obtain a second target feature map set corresponding to each second target; and the first target feature map set and the second target feature map set are fused to obtain a feature map set.

[0084] The first target feature map set contains first structured information corresponding to each first target, and the second target feature map set contains second structured information corresponding to each second target.

[0085] To facilitate intuitive distinction between the first target and the second target, the first target can be referred to as a large-size target, simply referred to as a large target; and the second target can be referred to as a small-size target, simply referred to as a small target. The large target can include a vehicle, a lane line, etc. The small target can include a traffic light, a traffic sign light. In the embodiments of the present application, a vehicle and a traffic light are taken as examples for introduction.

[0086] To improve the detection accuracy of the small target, two target detection heads in parallel can be used to extract the structured information of the small-size target and the large-size target.

[0087] The two detection heads used by the present application to extract the structured information of the small target and the large target can be referred to as a small target detection head (Head_Small) and a large target detection head (Head_Large), respectively. The Head_Large is used to detect a target object with a larger size, and the Head_Small is used to detect a small target with fine granularity and weak texture.

[0088] It should be noted that the BEV feature map can have different hierarchical structure forms in different model architectures. In a backbone structure with multi-scale BEV output capability, the above two detection heads can access different resolution feature layers from the BEV encoder, the Head_Large can be connected to low-resolution feature layers such as the fourth layer and the fifth layer, has a larger receptive field, and is suitable for perceiving large-size targets. The Head_Small is connected to high-resolution feature layers such as the second layer and the third layer, maintains spatial details, and is beneficial to small target modeling.

[0089] The detection head is a neural network module deployed in a software model, which is deployed in a BEV perception system, receives a BEV feature map as input, and uses multiple branch networks to extract structured information of a target of a specific size. To facilitate distinction, the branch network corresponding to the large target detection head can be referred to as a first branch network, and the branch network corresponding to the small target detection head can be referred to as a second branch network.

[0090] The first target quantity contained in the bird's eye view feature map is multiple, each first target has its corresponding first structured information, and the set of all first structured information is referred to as a first target feature map set; the second target quantity contained in the bird's eye view feature map is multiple, each second target has its corresponding second structured information, and the set of all second structured information is referred to as a second target feature map set.

[0091] In actual application, there may be a situation of conflict between the output results of the large target detection head and the small target detection head, that is, there is a region overlap between the first target detected by the large target detection head and the second target detected by the small target detection head.

[0092] In the embodiments of the present application, a region attention suppression mechanism is introduced to handle the conflict between the output results of the large target detection head and the small target detection head.

[0093] In the spatial dimension, the system takes the center position of each detection target as the center of a circle, constructs a target region according to a set radius, and uses it to detect the spatial conflict between the first target and the second target. When the center of the predicted target of the two detection heads is detected to fall into a certain overlapping region at the same time, the system performs differential processing according to the target category.

[0094] Taking a set target as a traffic light as an example, for the target falling into a certain overlapping region at the same time, the target is a traffic light category target, only the result from the small target detection head is retained. For non-traffic light category targets, a Soft-NMS (Soft-Non-Maximum Suppression) algorithm is introduced to perform confidence weighted fusion, effectively suppresses redundant boxes by decaying confidence, reduces the misjudgment of small targets, and at the same time avoids the misdeletion of effective targets in a multi-target proximity scene, and is particularly suitable for dense target detection scenes, such as traffic light or vehicle gathering areas.

[0095] If the two target regions have no spatial overlap, the first target feature map set and the second target feature map set are directly combined as a feature map set.

[0096] In a complex intersection scene, especially in the case of multiple traffic lights coexisting, some traffic lights being blocked, or the directions being inconsistent, the current traffic light recognition method often has problems such as inaccurate target recognition, missed detection, and false detection. Especially for the detection of distant and small target traffic lights, the current method lacks a special optimization mechanism, resulting in low detection accuracy, which affects the accuracy of subsequent decision-making and planning. The present application realizes accurate small target detection by designing two parallel detection heads and constructing independent traffic light channels in the BEV feature map, combining the height, direction angle, and relative distance to the ego vehicle of the traffic light, and improves the robustness and accuracy of target traffic light recognition.

[0097] S202: According to the target type to which each second target in the feature map set belongs, a first candidate set belonging to a set target is filtered out from the feature map set.

[0098] The first candidate set contains structured information of the target in the feature map set.

[0099] Taking a traffic light as an example, the structured information corresponding to the second target of the traffic light in the feature map set is taken as the first candidate set.

[0100] For example, the current ego coordinate is , and the ego heading angle is . The selected traffic light needs to serve motor vehicles and can be identified from the traffic light type label obtained from the BEV feature map or sensor supplemented structure, ;

[0101] wherein, represents a traffic light category judgment function, which is used to judge whether the traffic light I i belongs to a “motor vehicle light”.

[0102] S203: According to the spatial relationship between each second target and the ego, a second candidate set belonging to the target is selected from the feature map set.

[0103] The spatial relationship can include a positional relationship and an angular relationship.

[0104] In a specific implementation, the positional relationship between each second target in the feature map set and the ego driving area can be used to select a regional candidate set of the second target belonging to the ego driving area from the feature map set. The angular relationship between each second target in the feature map set and the ego can be used to select a directional candidate set of the second target matching the direction of the ego from the feature map set.

[0105] For the selection of the regional candidate set, the angle between each second target and the ego can be determined according to the deviation of the spatial position of each second target in the feature map set from the ego position and the unit vector of the forward direction of the ego; and the feature map set corresponding to a set number of second targets with the smallest angle with the ego in the feature map set is taken as the regional candidate set.

[0106] The regional candidate set contains structured information of the traffic light within the ego driving area.

[0107] In actual applications, the selected traffic light must fall within the near-field region of the forward driving path of the ego. By projecting the traffic light position vector onto the unit vector of the forward direction of the ego , if the spatial position of the traffic light is , the position of the ego is , and the unit vector of the forward direction of the ego is , then ; wherein, represents the angle between the traffic light and the ego.

[0108] According to The traffic lights are sorted by their values. A positive value indicates that the light is in front of the vehicle. The smaller the value, the closer the traffic light is to the vehicle.

[0109] In practical applications, all can be retained. And the k lights with the smallest values, i.e. ;

[0110] in, This function determines whether the projected traffic light is in front of the vehicle; a positive value indicates that the light is in front of the vehicle.

[0111] For the selection of direction candidate sets, the direction difference between each second target and the vehicle can be determined based on the orientation angle of each second target in the feature map set and the current direction of the vehicle. The feature map set corresponding to the second target in the feature map set whose direction difference is less than the set angle threshold is used as the direction candidate set.

[0112] The direction candidate set contains structured information about traffic lights that are in the same direction as the vehicle's current direction or whose direction difference is less than a set angle threshold.

[0113] In practical applications, the directional angle of traffic lights Should be in the current direction of the vehicle By maintaining a consistent angle or ensuring the angle difference does not exceed a set threshold, traffic lights facing other road traffic directions can be removed. The difference between the traffic light direction and the vehicle's direction is defined as: ;in, This represents the difference in direction between the i-th traffic light and the vehicle.

[0114] ;

[0115] in, A function to determine whether the direction of a traffic light is consistent with the direction of a vehicle. For the set angle threshold, satisfy This means the traffic light is considered to be aligned with the direction of the vehicle. Considering that in some road conditions the traffic light is not directly aligned with the stop line at the intersection, it can be... Set as .

[0116] S204: Normalize the first and second candidate sets to determine the rule scores corresponding to each second objective.

[0117] The second candidate set includes a region candidate set and a direction candidate set. In a specific implementation, the type candidate set, region candidate set, and direction candidate set can be normalized according to the following formula: ;in, This represents the rule score corresponding to the second objective.

[0118] In the embodiments of the present application, by quantifying the three types of rules into weight-adjustable scoring functions, the ranking values are output instead of discrete labels, the scoring functions retain ranking information compared with traditional binary rule judgment, which facilitates the learning module to access; weakly supervised pseudo label construction is supported, and gradient propagation between rules and deep learning models is supported.

[0119] S205: The feature map set corresponding to the second target with the highest rule score of a set number is taken as a target candidate set.

[0120] In actual application, the top k traffic lights can be retained as the target candidate set, and the target candidate set is denoted as , and the set is taken as the input basis of the subsequent module.

[0121] In the embodiments of the present application, in order to balance the accuracy and real-time performance of large target and small target detection in a complex environment, two detection heads in parallel are used to extract structured information of small size targets and large size targets from the bird's eye view feature map, independent modeling of branches for different scales of targets is realized, and the overall detection performance and hardware execution efficiency of the system are improved. Moreover, based on the rules, irrelevant light entities in the current decision of the ego vehicle are excluded, the search space is reduced, and the real-time performance and interpretability of the system are improved.

[0122] After obtaining the trained fusion classifier, if a new feature map set is obtained, a new target candidate set can be selected from the new feature map set according to the set target screening rule. The feature information of each new candidate target in the new target candidate set is semantically aligned with the ego vehicle intention vector to obtain a target fusion feature vector. The trained fusion classifier is used to analyze the target fusion feature vector to determine the probability that each new candidate target belongs to the set target; and the new candidate target with the highest probability is taken as the target associated with the ego vehicle.

[0123] Figure 3 ​A flowchart of a target traffic light selection process is provided for the embodiments of the present application, the target traffic light refers to the traffic light that needs to be followed during the driving of the ego vehicle. According to the operations involved in the selection of the target traffic light, the entire implementation process can be divided into different functional modules, including a parallel target detection head, a rule screening module, an intention guiding module, a fusion classifier, a trajectory prediction, and a trajectory encoder. The parallel target detection head is used to extract the structured information of small targets and large targets from the bird's eye view feature map respectively. In order to avoid category conflict or misjudgment caused by multi-head output, the parallel target detection head can fuse the first target feature map set and the second target feature map set, thereby obtaining the feature map set. The rule screening module is used to exclude the traffic lights in the feature map set that are irrelevant to the ego vehicle, thereby determining the candidate traffic light set, which contains each candidate traffic light and its corresponding rule score. The trajectory prediction is used to extract the trajectory point sequence of the ego vehicle. The trajectory encoder is used to encode the trajectory point sequence of the ego vehicle into the intention vector of the ego vehicle. The intention guiding module is used to semantically align the structured information of the candidate traffic light with the intention vector of the ego vehicle, thereby obtaining the fusion feature vector. The fusion classifier is used to analyze the probability of each candidate traffic light belonging to the target traffic light in the fusion feature vector, and the candidate traffic light with the highest probability is taken as the target traffic light associated with the ego vehicle. The output result of the fusion classifier will affect the driving path of the ego vehicle, thereby generating a new trajectory point sequence. In the training phase of the fusion classifier, a pseudo-label generation module is included, which is used to generate pseudo-labels based on the candidate light rule score and the intention matching score, and the weak supervision training based on the pseudo-labels realizes the training of the fusion classifier. The trained fusion classifier can accurately identify the target traffic light.

[0124] The target recognition scheme proposed in the present application can be directly deployed in edge computing series products. The model deployment adopts a modular deployment strategy: including perception backbone (BEV encoder), small target detection head, rule screening module, intention guiding module, etc. structural units, graph fusion and quantization processing are performed through end-side inference optimization tools, the model volume is compressed and the inference efficiency is improved. During deployment, the model receives image input from multiple vehicle-mounted cameras, generates BEV feature map after front-end perception processing, and then extracts small target information and judges the target traffic light through the structure proposed in the present application. The state and position of the target traffic light are written as structured results in the edge cache for real-time reading by downstream path planning and control modules. The overall inference delay of the system is controlled within 50ms, meeting the perception response requirements of system edge deployment.

[0125] Figure 4 A structural diagram of a target recognition device is provided for the embodiments of the present application, including a screening unit 41, an alignment unit 42, a matching unit 43, a construction unit 44, and an adjustment unit 45.

[0126] The screening unit 41 is configured to screen target candidate sets from the feature map set according to a set target screening rule; the target screening rule is set according to the relevance between the set target and the driving behavior of the ego vehicle; and each candidate target and a corresponding rule score are included in the target candidate set.

[0127] The alignment unit 42 is configured to perform semantic alignment on the feature information of each candidate target in the target candidate set and the ego vehicle intention vector to obtain a fusion feature vector.

[0128] The matching unit 43 is configured to analyze the visual representation features of each candidate target and the ego vehicle intention vector to obtain an intention matching score of each candidate target.

[0129] The construction unit 44 is configured to construct pseudo-label data based on the intention matching score of each candidate target and the rule score; and the pseudo-label data includes a predicted probability that each candidate target belongs to the set target.

[0130] The adjustment unit 45 is configured to adjust the parameters of the fusion classifier according to the pseudo-label data and the recognition result output by analyzing the fusion feature vector by the fusion classifier, so as to obtain a trained fusion classifier, and to identify the set target by using the trained fusion classifier.

[0131] In some embodiments, the screening unit includes an extraction subunit, a type matching subunit, a spatial relationship matching subunit, a normalization subunit, and a selection subunit.

[0132] The extraction subunit is configured to extract a feature map set from a bird's eye view feature map by using a parallel branch network; the bird's eye view feature map is a feature map obtained by encoding a bird's eye view image by using a bird's eye view encoding method; the feature map set includes a feature map set corresponding to a first target and a feature map set corresponding to a second target; and the number of pixels of the first target is greater than the number of pixels of the second target.

[0133] The type matching subunit is configured to screen a first candidate set belonging to the set target from the feature map set according to the target type to which each second target in the feature map set belongs.

[0134] The spatial relationship matching subunit is configured to screen a second candidate set belonging to the set target from the feature map set according to the spatial relationship between each second target and the ego vehicle.

[0135] The normalization subunit is configured to perform normalization processing on the first candidate set and the second candidate set to determine a rule score corresponding to each second target.

[0136] The selection subunit is configured to select the feature map set corresponding to a set number of second targets with the highest rule scores as the target candidate set.

[0137] In some embodiments, the extraction subunit is configured to perform feature extraction on the bird's eye view feature map using the first branch network to obtain a first target feature map set corresponding to each first target;

[0138] perform feature extraction on the bird's eye view feature map using the second branch network to obtain a second target feature map set corresponding to each second target;

[0139] fuse the first target feature map set and the second target feature map set to obtain a feature map set.

[0140] In some embodiments, the spatial relationship matching subunit is configured to filter out a region candidate set of the second targets belonging to the ego vehicle driving area from the feature map set according to the positional relationship between each second target in the feature map set and the ego vehicle driving area;

[0141] filter out a direction candidate set of the second targets matching the direction of the ego vehicle from the feature map set according to the angle relationship between each second target in the feature map set and the ego vehicle.

[0142] In some embodiments, the spatial relationship matching subunit is configured to determine the angle between each second target and the ego vehicle according to the deviation of the spatial position of each second target in the feature map set from the position of the ego vehicle and the unit vector of the forward direction of the ego vehicle;

[0143] take the feature map set corresponding to a set number of second targets with the smallest angle with the ego vehicle in the feature map set as the region candidate set.

[0144] In some embodiments, the spatial relationship matching subunit is configured to determine the direction difference between each second target and the ego vehicle according to the orientation angle of each second target in the feature map set and the current direction of the ego vehicle;

[0145] take the feature map set corresponding to the second targets with a direction difference less than a set angle threshold in the feature map set as the direction candidate set.

[0146] In some embodiments, the alignment unit includes an encoding subunit, an extraction subunit, a first fusion subunit, and a second fusion subunit;

[0147] The encoding subunit is configured to encode the sequence of trajectory points of the ego vehicle into an ego vehicle intention vector;

[0148] The extraction subunit is configured to extract feature information of each candidate target from the target candidate set; wherein the feature information includes visual representation features, direction angle cosine values, and relative distances between the candidate targets and the ego vehicle;

[0149] The first fusion subunit is configured to fuse the ego vehicle intention vector and the feature information of each candidate target to obtain an intention vector of each candidate target;

[0150] The second fusion subunit is configured to align the intention vector of each candidate target with the ego intention vector according to the key-value pair information of each position in the bird's eye view feature map, so as to obtain a fusion feature vector.

[0151] In some embodiments, the matching unit comprises a splicing subunit and an analysis subunit.

[0152] The splicing subunit is configured to splice the visual representation feature of each candidate target and the ego intention vector, so as to obtain a spliced vector.

[0153] The analysis subunit is configured to analyze the spliced vector by using a multi-layer fully connected layer, and map the analysis result as an intention matching score.

[0154] In some embodiments, the construction unit is configured to perform probability normalization processing on the intention matching score of each candidate target and the rule score, so as to obtain a predicted probability that each candidate target belongs to the set target.

[0155] In some embodiments, the adjustment unit is configured to input the fusion feature vector into the fusion classifier, so as to determine a probability that each candidate target belongs to the set target.

[0156] According to the probability that each candidate target belongs to the set target and the predicted probability that each candidate target belongs to the set target, a cross-entropy loss function is constructed.

[0157] The parameters of the fusion classifier are adjusted by using the cross-entropy loss function, so as to obtain a trained fusion classifier.

[0158] In some embodiments, the target recognition device further comprises an analysis unit and an as unit.

[0159] The screening unit is configured to, in a case where a new feature map set is obtained, screen a new target candidate set from the new feature map set according to a set target screening rule.

[0160] The alignment unit is configured to perform semantic alignment between the feature information of each new candidate target in the new target candidate set and the ego intention vector, so as to obtain a target fusion feature vector.

[0161] The analysis unit is configured to analyze the target fusion feature vector by using the trained fusion classifier, so as to determine a probability that each new candidate target belongs to the set target.

[0162] The as unit is configured to take the new candidate target with the largest probability as the target associated with the ego vehicle.

[0163] The features of the embodiments corresponding to the target recognition device can be referred to the related descriptions of the embodiments corresponding to the target recognition method, which will not be repeated here.

[0164] According to the technical solution, the target candidate set is filtered from the feature set according to the set target filtering rule, wherein the target filtering rule is set according to the relevance between the set target and the driving behavior of the ego vehicle, and the target candidate set includes each candidate target and a corresponding rule score. By setting the target filtering rule, irrelevant targets in the feature set can be excluded. To increase the correctness of target selection, the feature information of each candidate target in the target candidate set can be semantically aligned with the ego vehicle intention vector to obtain a fusion feature vector. The visual representation feature of each candidate target and the ego vehicle intention vector are analyzed to obtain an intention matching score of each candidate target. Based on the intention matching score and the rule score of each candidate target, pseudo-label data is constructed, wherein the pseudo-label data includes a predicted probability of each candidate target belonging to the set target. According to the pseudo-label data and the recognition result output by analyzing the fusion feature vector by the fusion classifier, the parameters of the fusion classifier are adjusted to obtain a trained fusion classifier, so as to identify the set target by using the trained fusion classifier. In this application, based on the filtered target candidate set, a semantic alignment mechanism between the ego vehicle intention and the candidate target is established by intention guidance to solve the problem of distinguishing multiple candidate targets with different priorities in the same scene. Based on the weak supervision training mechanism of the pseudo-label, the rule score and the intention matching score of the candidate target are fused to generate a pseudo-label distribution, and the weak supervision training of the fusion classifier is performed to solve the problem of difficult construction of the data set and scarcity of real labels in the target detection and recognition task. Moreover, the weak supervision training mechanism based on the pseudo-label improves the generalization ability of the fusion classifier in a complex environment. The trained fusion classifier can accurately identify the set target.

[0165] The embodiment of the present application further provides an electronic device, including a memory and a processor, the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any one of the above target identification method embodiments.

[0166] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program, wherein the computer program is configured to execute the steps in any one of the above target identification method embodiments when running.

[0167] In an example embodiment, the above computer readable storage medium can include, but is not limited to, a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store computer programs.

[0168] The embodiment of the present application further provides a computer program product, which comprises a computer program, and the computer program realizes the steps in any of the target recognition method embodiments when executed by a processor.

[0169] The embodiment of the present application further provides another computer program product, which comprises a nonvolatile computer readable storage medium, and the nonvolatile computer readable storage medium stores a computer program, and the computer program realizes the steps in any of the target recognition method embodiments when executed by a processor.

[0170] Those skilled in the art can further understand that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been described in general in the above description. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0171] The above describes in detail the target recognition method, device, equipment, storage medium and product provided by the present application. The principles and implementation modes of the present application are described by applying specific examples in this paper. The above description of the examples is only applicable to help understand the method and core idea of the present application. It should be pointed out that for those skilled in the art, without departing from the principles of the present application, some improvements and modifications can be made to the present application, and these improvements and modifications also fall within the protection scope of the present application.

Claims

1. A target recognition method characterized by, The method comprises the following steps: According to the set target screening rule, a target candidate set is screened out from the feature map set; wherein, the target screening rule is set according to the relevance between the set target and the driving behavior of the ego vehicle; the target candidate set contains each candidate target and its corresponding rule score; The semantic alignment is performed between the feature information of each candidate target in the target candidate set and the ego vehicle intention vector to obtain a fusion feature vector; The visual representation features of each candidate target and the ego vehicle intention vector are analyzed to obtain an intention matching score of each candidate target; Based on the intention matching score and the rule score of each candidate target, pseudo-label data is constructed; wherein, the pseudo-label data contains the predicted probability of each candidate target belonging to the set target; According to the pseudo-label data and the recognition result output by the analysis of the fusion classifier on the fusion feature vector, the parameters of the fusion classifier are adjusted to obtain a trained fusion classifier, so as to identify the set target by using the trained fusion classifier; According to the set target screening rule, a target candidate set is screened out from the feature map set, which comprises the following steps: A feature map set is extracted from the bird's eye view feature map by using a parallel branch network; wherein, the bird's eye view feature map is a feature map obtained by encoding the multi-camera image into a bird's eye view image; the feature map set includes a feature map set corresponding to a first target and a feature map set corresponding to a second target, and the number of pixels of the first target is greater than that of the second target; According to the target type to which each second target in the feature map set belongs, a first candidate set belonging to the set target is screened out from the feature map set; According to the spatial relationship between each second target and the ego vehicle, a second candidate set belonging to the set target is screened out from the feature map set; the spatial relationship includes the position relationship and the angle relationship; the second candidate set includes a region candidate set and a direction candidate set; the region candidate set contains the structured information of the traffic light within the driving region of the ego vehicle; the direction candidate set contains the structured information of the traffic light with the same direction as the current direction of the ego vehicle or the direction difference less than a set angle threshold; The first candidate set and the second candidate set are normalized to determine the rule score corresponding to each second target; The feature map set corresponding to the set number of second targets with the highest rule score is taken as the target candidate set.

2. The object recognition method of claim 1, wherein, The feature map set is extracted from the bird's eye view feature map by using a parallel branch network, which comprises the following steps: The first branch network is used for feature extraction on the bird's eye view feature map to obtain a first target feature map set corresponding to each first target; The second branch network is used for feature extraction on the bird's eye view feature map to obtain a second target feature map set corresponding to each second target; The first target feature map set and the second target feature map set are fused to obtain the feature map set.

3. The object recognition method of claim 1, wherein, According to the spatial relationship between each second target and the ego vehicle, a second candidate set belonging to the set target is screened out from the feature map set, which comprises the following steps: The position relationship between each second target in the feature map set and the driving region of the ego vehicle is used to screen out a region candidate set of the second target belonging to the driving region of the ego vehicle from the feature map set; Screening, from the feature map set, a direction candidate set of the second targets matching a direction of the ego vehicle according to an angle relationship between each of the second targets and the ego vehicle in the feature map set.

4. The object recognition method of claim 3, wherein, Screening, from the feature map set, a region candidate set of the second targets belonging to a driving region of the ego vehicle according to a position relationship between each of the second targets and the driving region of the ego vehicle in the feature map set, comprising: Determining an angle between each of the second targets and the ego vehicle according to a deviation of a spatial position of each of the second targets from a position of the ego vehicle and a unit vector of a forward direction of the ego vehicle; Taking, as the region candidate set, a feature map set corresponding to a set number of the second targets with the smallest angles to the ego vehicle in the feature map set.

5. The object recognition method of claim 3, wherein, Screening, from the feature map set, a direction candidate set of the second targets matching a direction of the ego vehicle according to an angle relationship between each of the second targets and the ego vehicle in the feature map set, comprising: Determining a direction difference between each of the second targets and the ego vehicle according to an orientation angle of each of the second targets and a current direction of the ego vehicle in the feature map set; Taking, as the direction candidate set, a feature map set corresponding to the second targets with a direction difference less than a set angle threshold in the feature map set.

6. The object recognition method of claim 1, wherein, Aligning, semantically, feature information of each of the candidate targets in the target candidate set with an ego intention vector to obtain a fusion feature vector, comprising: Encoding a trajectory point sequence of the ego vehicle into the ego intention vector; Extracting, from the target candidate set, the feature information of each of the candidate targets, wherein the feature information comprises visual representation features, a direction angle cosine value, and a relative distance between the candidate target and the ego vehicle; Fusing the ego intention vector with the feature information of each of the candidate targets to obtain an intention vector of each of the candidate targets; Aligning, according to key-value pair information of each position in an overhead feature map, the intention vector of each of the candidate targets with the ego intention vector to obtain a fusion feature vector.

7. The object recognition method of claim 1, wherein, Analyzing the visual representation features of each of the candidate targets and the ego intention vector to obtain an intention matching score of each of the candidate targets, comprising: Splicing the visual representation features of each of the candidate targets and the ego intention vector to obtain a spliced vector; Analyzing the spliced vector by using a plurality of fully connected layers and mapping an analysis result into the intention matching score.

8. The object recognition method of claim 7, wherein, Constructing pseudo-label data based on the intention matching score of each of the candidate targets and a rule score, comprising: Performing probability normalization processing on the intention matching score of each of the candidate targets and the rule score to obtain a prediction probability of each of the candidate targets belonging to a set target.

9. The object recognition method of claim 8, wherein, Adjusting parameters of a fusion classifier according to an identification result output by the fusion classifier when analyzing the fusion feature vector, to obtain a trained fusion classifier, comprising: Inputting the fusion feature vector into the fusion classifier to determine a probability of each of the candidate targets belonging to a set target; Constructing a cross-entropy loss function according to the probability of each of the candidate targets belonging to the set target and a prediction probability of each of the candidate targets belonging to the set target. The parameters of the fusion classifier are adjusted by using the cross-entropy loss function to obtain a trained fusion classifier.

10. The object recognition method according to any one of claims 1 to 9, characterized in that, After obtaining the trained fusion classifier, further comprising: In the case of obtaining a new feature map set, filtering a new target candidate set from the new feature map set according to a set target filtering rule; Aligning the feature information of each new candidate target in the new target candidate set with the ego vehicle intention vector to obtain a target fusion feature vector; Using the trained fusion classifier to analyze the target fusion feature vector to determine the probability of each new candidate target belonging to the set target; The new candidate target with the highest probability is taken as the target associated with the ego vehicle.

11. A target recognition device, characterized by The method comprises a filtering unit, an alignment unit, a matching unit, a construction unit and an adjustment unit; The filtering unit is configured to filter a target candidate set from a feature map set according to a set target filtering rule; wherein the target filtering rule is set according to the relevance between the set target and the ego vehicle driving behavior; the target candidate set comprises each candidate target and its corresponding rule score; The alignment unit is configured to align the feature information of each candidate target in the target candidate set with the ego vehicle intention vector to obtain a fusion feature vector; The matching unit is configured to analyze the visual representation features of each candidate target and the ego vehicle intention vector to obtain an intention matching score of each candidate target; The construction unit is configured to construct pseudo-label data based on the intention matching score and the rule score of each candidate target; wherein the pseudo-label data comprises the predicted probability of each candidate target belonging to the set target; The adjustment unit is configured to adjust the parameters of the fusion classifier according to the pseudo-label data and the recognition result output by the fusion classifier when analyzing the fusion feature vector, to obtain a trained fusion classifier, so as to identify the set target by using the trained fusion classifier. The screening unit comprises an extraction subunit, a type matching subunit, a spatial relationship matching subunit, a normalization subunit and an as subunit; the extraction subunit is configured to extract a feature map set from a bird's eye view feature map by using a parallel branch network; the bird's eye view feature map is a feature map obtained by encoding a bird's eye view map from multiple camera images; the feature map set comprises a feature map set corresponding to a first target and a feature map set corresponding to a second target, the number of pixel points of the first target is greater than the number of pixel points of the second target; the type matching subunit is configured to screen a first candidate set belonging to a set target from the feature map set according to a target type to which each second target belongs; the spatial relationship matching subunit is configured to screen a second candidate set belonging to the set target from the feature map set according to a spatial relationship of each second target with the ego vehicle; the spatial relationship comprises a position relationship and an angle relationship; the second candidate set comprises a region candidate set and a direction candidate set; the region candidate set comprises structured information of traffic lights within a driving region range of the ego vehicle; the direction candidate set comprises structured information of traffic lights consistent with a current direction of the ego vehicle or having a direction difference less than a set angle threshold; the normalization subunit is configured to perform normalization processing on the first candidate set and the second candidate set to determine a rule score corresponding to each second target; and the as subunit is configured to take the feature map set corresponding to a set number of second targets with the highest rule scores as a target candidate set.

12. An electronic device, comprising: The memory is configured to store a computer program; The processor is configured to implement the steps of the target identification method according to any one of claims 1 to 10 when executing the computer program. The computer readable storage medium stores a computer program, and the computer program is configured to implement the steps of the target identification method according to any one of claims 1 to 10 when executed by the processor.

13. A computer-readable storage medium, characterized in that, The computer program is configured to implement the steps of the target identification method according to any one of claims 1 to 10 when executed by the processor.

14. A computer program product comprising a computer program, characterized in that, ​

Images