Target detection method and device, computer equipment and storage medium

Abstract

The invention relates to the technical field of computers, and discloses a target detection method and device, computer equipment and a storage medium, and the method comprises the steps: inputting a target image into a target detection network, and obtaining prediction score information of a plurality of first candidate boxes; screening out a second candidate frame from the plurality of first candidate frames according to the target category prediction score and existence prediction score of each first candidate frame, a category prediction score threshold value and an existence prediction score threshold value; for each second candidate frame, converting the target category prediction score of the second candidate frame into a category prediction probability of the second candidate frame, and determining whether the second candidate frame is used as a third candidate frame according to the category prediction probability of the second candidate frame and a category prediction probability threshold; and determining a target detection result corresponding to the target image according to each third candidate box.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a target detection method, device, computer equipment and storage medium. Background Art

[0002] Detecting targets around the vehicle through a target detection network is a basic function of systems such as assisted driving systems and autonomous driving systems. In related technologies, when post-processing the candidate box output by the target detection network, the two scores of the candidate box are converted into corresponding probabilities using the Sigmoid() function, and the product of the two probabilities is used as the confidence of the candidate box. The final confidence of the candidate box is compared with a fixed confidence threshold to determine whether the candidate box is used as a candidate box for the algorithm such as NMS for generating the detection box of the target detection result corresponding to the target image.

[0003] However, using the Sigmoid() function to convert the two predicted scores of the candidate box into corresponding probabilities and taking the product of the two probabilities as the confidence of the candidate box has high computational complexity. The number of candidate boxes generated by the target detection network is large, and all candidate boxes output by the target detection network need to be operated with high computational complexity, resulting in low efficiency of target detection. How to improve the efficiency of target detection has become a problem that needs to be solved. Summary of the invention

[0004] In view of this, embodiments of the present application provide a target detection method, apparatus, computer device, and storage medium.

[0005] In a first aspect, an embodiment of the present application provides a target detection method, comprising:

[0006] Inputting the target image into the target detection network to obtain prediction score information of multiple first candidate boxes, the prediction score information of the first candidate boxes including: a category prediction score and an existence prediction score for each category in the multiple categories;

[0007] According to the target category prediction score and existence prediction score of each first candidate box, the category prediction score threshold, and the existence prediction score threshold, a second candidate box is screened out from the multiple first candidate boxes, wherein the target category prediction score of the first candidate box is the largest score among the multiple category prediction scores of the first candidate boxes;

[0008] For each second candidate box, converting the target category prediction score of the second candidate box into the category prediction probability of the second candidate box, and determining whether the second candidate box is used as the third candidate box according to the category prediction probability of the second candidate box and the category prediction probability threshold;

[0009] According to each third candidate box, a target detection result corresponding to the target image is determined.

[0010] In a possible implementation, screening out a second candidate box from a plurality of first candidate boxes according to a target category prediction score and an existence prediction score of each first candidate box, a category prediction score threshold, and an existence prediction score threshold includes:

[0011] For each first candidate box, when the existence prediction score of the first candidate box is greater than the existence prediction score threshold and the target category prediction score of the first candidate box is greater than the category prediction score threshold, the first candidate box is used as the second candidate box.

[0012] In a possible implementation, determining whether the second candidate box is used as the third candidate box includes, according to the category prediction probability and the category prediction probability threshold of the second candidate box:

[0013] Determine, according to the association information of the second candidate box, a category prediction probability threshold corresponding to the second candidate box, the association information including at least some of the following items: a target depth corresponding to the second candidate box, a target category corresponding to the second candidate box, and a scale of a target feature map corresponding to the second candidate box, the target category is a category for which the target category prediction score of the second candidate box is targeted, and the target feature map is a feature map used when predicting the target category prediction score of the second candidate box;

[0014] According to the category prediction probability of the second candidate box and the category prediction probability threshold corresponding to the second candidate box, it is determined whether the second candidate box is used as the third candidate box.

[0015] In a possible implementation, determining, according to the association information of the second candidate box, a category prediction probability threshold corresponding to the second candidate box includes:

[0016] When the target category corresponding to the second candidate box is a vehicle category, determining a category prediction probability threshold corresponding to the second candidate box according to the scale of the target feature map corresponding to the second candidate box;

[0017] When the target category corresponding to the second candidate box is an obstacle category or a vulnerable road user category, a category prediction probability threshold corresponding to the second candidate box is determined according to the target depth corresponding to the second candidate box.

[0018] In a possible implementation, determining the category prediction probability threshold corresponding to the second candidate box according to the scale of the target feature map corresponding to the second candidate box includes:

[0019] According to the scale of the target feature map corresponding to the second candidate box and the feature map scale weighting coefficient, the category prediction probability threshold corresponding to the second candidate box is determined.

[0020] In a possible implementation, determining, according to the target depth corresponding to the second candidate box, a category prediction probability threshold corresponding to the second candidate box includes:

[0021] According to the target depth, the depth weighting coefficient, and the category weighting coefficient corresponding to the second candidate box, a category prediction probability threshold corresponding to the second candidate box is determined.

[0022] In a second aspect, an embodiment of the present application provides a target detection device installed on a vehicle, the target detection device comprising:

[0023] An input unit, used to input the target image into the target detection network to obtain prediction score information of multiple first candidate boxes, wherein the prediction score information of the first candidate boxes includes: a category prediction score and an existence prediction score for each category in the multiple categories;

[0024] A first screening unit is used to screen out a second candidate frame from multiple first candidate frames according to a target category prediction score and an existence prediction score of each first candidate frame, a category prediction score threshold, and an existence prediction score threshold, wherein the target category prediction score of the first candidate frame is the largest score among the multiple category prediction scores of the first candidate frames;

[0025] A second screening unit is used to convert the target category prediction score of the second candidate box into a category prediction probability of the second candidate box for each second candidate box, and determine whether the second candidate box is used as the third candidate box according to the category prediction probability of the second candidate box and the category prediction probability threshold;

[0026] The target detection result determination unit is used to determine the target detection result corresponding to the target image according to each third candidate frame.

[0027] In one possible implementation, the first screening unit is further used to, for each first candidate box, use the first candidate box as the second candidate box when the existence prediction score of the first candidate box is greater than the existence prediction score threshold and the target category prediction score of the first candidate box is greater than the category prediction score threshold.

[0028] In a possible implementation, the second screening unit is also used to determine a category prediction probability threshold corresponding to the second candidate box based on the association information of the second candidate box, the association information including at least some of the following items: a target depth corresponding to the second candidate box, a target category corresponding to the second candidate box, and a scale of a target feature map corresponding to the second candidate box, the target category is a category for which the target category prediction score of the second candidate box is targeted, and the target feature map is a feature map used when predicting the target category prediction score of the second candidate box; determine whether the second candidate box is used as the third candidate box based on the category prediction probability of the second candidate box and the category prediction probability threshold corresponding to the second candidate box.

[0029] In a possible implementation, the second screening unit is also used to determine the category prediction probability threshold corresponding to the second candidate box according to the scale of the target feature map corresponding to the second candidate box when the target category corresponding to the second candidate box is a vehicle category; when the target category corresponding to the second candidate box is an obstacle category or a vulnerable road user category, determine the category prediction probability threshold corresponding to the second candidate box according to the target depth corresponding to the second candidate box.

[0030] In a possible implementation, the second screening unit is also used to determine the category prediction probability threshold corresponding to the second candidate box according to the scale of the target feature map corresponding to the second candidate box and the feature map scale weighting coefficient when the target category corresponding to the second candidate box is a vehicle category.

[0031] In one possible implementation, the second screening unit is also used to determine the category prediction probability threshold corresponding to the second candidate box based on the target depth, depth weighting coefficient, and category weighting coefficient corresponding to the second candidate box when the target category corresponding to the second candidate box is an obstacle category or a vulnerable road user category.

[0032] In a third aspect, an embodiment of the present application provides a computer device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, computer instructions being stored in the memory, and the processor executing the method of the first aspect or any corresponding embodiment thereof by executing the computer instructions.

[0033] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having computer instructions stored thereon, the computer instructions being used to enable a computer to execute the method of the first aspect or any corresponding embodiment thereof.

[0034] In a fifth aspect, the present invention provides a computer program product, comprising computer instructions for causing a computer to execute the method of the first aspect or any corresponding embodiment thereof.

[0035] The target detection method provided in the embodiment of the present application screens out second candidate frames from multiple first candidate frames based on the target category prediction score and existence prediction score of each first candidate frame, the category prediction score threshold, and the existence prediction score threshold; for each second candidate frame, converts the target category prediction score of the second candidate frame into the category prediction probability of the second candidate frame, and determines whether the second candidate frame is used as the third candidate frame based on the category prediction probability of the second candidate frame and the category prediction probability threshold; and determines the target detection result corresponding to the target image based on each third candidate frame.

[0036] On the one hand, in the target detection method provided in an embodiment of the present application, second candidate boxes are screened out from multiple first candidate boxes, the number of second candidate boxes is less than the number of first candidate boxes, and the target category prediction score of each second candidate box is converted into the category prediction probability of the second candidate box. It is only necessary to convert the target category prediction scores of all candidate boxes output by the target detection network, that is, a part of all first candidate boxes, that is, the second candidate boxes, into the category prediction probability of the second candidate boxes.

[0037] Compared to the related art, in which all candidate boxes output by the target detection network are converted into corresponding probabilities respectively for the two scores of the candidate boxes, namely the existence prediction score and the category prediction score, the target detection method provided in the embodiment of the present application has a small number of candidate boxes involved in the conversion from score to probability, and at the same time, the number of scores of a single candidate box involved in the conversion from score to probability is small. Therefore, through the target method provided in the embodiment of the present application, the algorithm for determining the candidate box that may be used as the detection box for generating the target detection result corresponding to the target image, such as the NMS algorithm, is efficient, thereby improving the efficiency of target detection.

[0038] On the other hand, in the target detection method provided in an embodiment of the present application, for each second candidate box, it is determined based on the category prediction probability and the category prediction probability threshold of the second candidate box whether the second candidate box is used as the third candidate box, that is, the algorithm used to determine the detection box in the target detection result, for example, the candidate box targeted by the non-maximum suppression (NMS) algorithm.

[0039] The target detection method provided in the embodiment of the present application needs to determine whether the candidate box used as the third candidate box is only a part of all candidate boxes output by the target detection network, that is, a part of all first candidate boxes, that is, the second candidate box. The target detection method provided in the embodiment of the present application needs to determine whether the number of candidate boxes used as the third candidate box is less than the number of all candidate boxes output by the target detection network, that is, the number of all first candidate boxes.

[0040] Compared with the related art in which it is separately determined whether each candidate box among all candidate boxes output by the target detection network is used as a candidate box targeted by an algorithm such as the NMS algorithm for determining a detection box for generating a target detection result, the target detection method provided in the embodiment of the present application needs to determine whether the number of candidate boxes targeted by the algorithm is small, thereby improving the efficiency of target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the specific implementation methods of the present application or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] Figure 1 It is a flowchart of the target detection method provided in the embodiment of the present application;

[0043] Figure 2 is a flowchart of an example of post-processing of target detection by the target detection method provided by an embodiment of the present application;

[0044] Figure 3 is a flow chart of another target detection method provided in an embodiment of the present application;

[0045] Figure 4 is a flowchart of another example of post-processing of target detection by the target detection method provided by an embodiment of the present application;

[0046] Figure 5 It is a schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of this application.

[0048] refer to Figure 1 , which shows a flow chart of the target detection method provided in an embodiment of the present application.

[0049] In step S101, the target image is input into the target detection network to obtain the prediction score information of multiple first candidate boxes.

[0050] The target can be any image that contains the target.

[0051] As an example, the target image is an environmental image of the environment of the vehicle captured by a camera of the vehicle during driving of the vehicle.

[0052] Step S101: input a target image into a target detection network, and the target detection network outputs a plurality of first candidate boxes and prediction score information of the plurality of first candidate boxes.

[0053] The first candidate frame is a two-dimensional frame and can be represented by the coordinates of the center point of the first candidate frame, the length of the first candidate frame, and the width of the first candidate frame.

[0054] Typically, each first candidate frame in the partial first candidate frames surrounds a corresponding target or a part of the corresponding target in the target image.

[0055] The target detection network is trained, that is, before executing step S101, the target detection network is trained using a training set of the target detection network.

[0056] For a first candidate box, the prediction score information of the first candidate box includes: a category prediction score of the first candidate box for each category in a plurality of categories, and an existence prediction score of the first candidate box.

[0057] As an example, the multiple categories consist of category 1, category 2...category N. For a first candidate box, the prediction score information of the first candidate box includes: the category prediction score of the first candidate box for category 1, the category prediction score of the first candidate box for category 2...the category prediction score of the first candidate box for category N.

[0058] For a first candidate box and a category, the category prediction score of the first candidate box for the category is positively correlated with the probability that the target in the candidate box belongs to the category. The larger the category prediction score of the first candidate box for the category, the greater the probability that the target in the first candidate box belongs to the category. The smaller the category prediction score of the first candidate box for the category, the smaller the probability that the target in the first candidate box belongs to the category.

[0059] For a first candidate box, the existence prediction score of the first candidate box is positively correlated with the probability of the existence of the target in the first candidate box. The larger the existence prediction score of the first candidate box, the greater the probability of the existence of the target in the first candidate box. The smaller the existence prediction score of the first candidate box, the smaller the probability of the existence of the target in the first candidate box.

[0060] In a possible implementation, the target detection network used in step S101 is an anchor-free target detection network such as CornerNet. If the target detection network used in step S101 is an anchor-free target detection network, the existence prediction score of the candidate box can also be called the centerness prediction score.

[0061] In another possible implementation, the target detection network used in step S101 is an anchor-based target detection network. As an example, the target detection network used in step S101 is a YOLO series target detection network.

[0062] In step S102, second candidate boxes are screened out from a plurality of first candidate boxes according to the target category prediction score and existence prediction score of each first candidate box, the category prediction score threshold, and the existence prediction score threshold.

[0063] The number of the second candidate boxes is smaller than the number of the first candidate boxes.

[0064] For a first candidate box, the target category prediction score of the first candidate box is the largest score among the multiple category prediction scores of the first candidate box, wherein the multiple category prediction scores of the first candidate box correspond one-to-one to the multiple categories.

[0065] As an example, the above multiple categories consist of category 1, category 2...category N. For a first candidate box, the prediction score information of the first candidate box includes: the category prediction score of the first candidate box for category 1, the category prediction score of the first candidate box for category 2...the category prediction score of the first candidate box for category N. Among them, for category j among multiple categories, the category prediction score of the first candidate box for category j is greater than the score of the first candidate box for any other category among the multiple categories, and the category prediction score of the first candidate box for category j is: the target category prediction score of the first candidate box.

[0066] In step S102, for each first candidate box, it is determined whether the first candidate box is used as the second candidate box based on the comparison result of the target category prediction score of the first candidate box with the category prediction score threshold and the comparison result of the existence prediction score threshold of the first candidate box with the existence prediction score threshold.

[0067] For example, the target detection network outputs N first candidate boxes, including the first candidate box 1, the first candidate box 2, ... the first candidate box N. According to the comparison result of the target category prediction score of the first candidate box 1 and the category prediction score threshold, and the comparison result of the existence prediction score threshold of the first candidate box 1 and the existence prediction score threshold, it is determined whether the first candidate box 1 is used as the second candidate box. According to the comparison result of the target category prediction score of the first candidate box 2 and the category prediction score threshold, and the comparison result of the existence prediction score threshold of the first candidate box 2 and the existence prediction score threshold, it is determined whether the first candidate box 2 is used as the second candidate box. According to the comparison result of the target category prediction score of the first candidate box N and the category prediction score threshold, and the comparison result of the existence prediction score threshold of the first candidate box N and the existence prediction score threshold, it is determined whether the first candidate box N is used as the second candidate box.

[0068] In one possible implementation, in step S102, for a first candidate box, if the target category prediction score of the first candidate box is greater than the category prediction score threshold, the first candidate box can be used as the second candidate box. For a first candidate box, if the existence prediction score of the first candidate box is greater than the existence prediction score threshold, the first candidate box can be used as the second candidate box. In step S102, for a first candidate box, if the target category prediction score of the first candidate box is not greater than the category prediction score threshold and the existence prediction score of the first candidate box is not greater than the existence prediction score threshold, it is determined that the first candidate box is not used as the second candidate box.

[0069] In step S103, for each second candidate box, the target category prediction score of the second candidate box is converted into the category prediction probability of the second candidate box, and according to the category prediction probability of the second candidate box and the category prediction probability threshold, it is determined whether the second candidate box is used as the third candidate box.

[0070] Among them, for a second candidate box, the category prediction probability of the second candidate box can indicate: the probability that at least a part of the target surrounded by the second candidate box belongs to the category targeted by the target category prediction score of the second candidate box. It should be noted that the category targeted by the target category prediction score of the second candidate box is the target category corresponding to the second candidate box.

[0071] In step S103, any function for converting the score output by the target detection network into a probability may be used to convert the target category prediction score of the second candidate box into a category prediction probability of the second candidate box.

[0072] As an example, the target category prediction score of the second candidate box is converted into the category prediction probability of the second candidate box using the sigmoid function. As another example, the target category prediction score of the second candidate box is converted into the category prediction probability of the second candidate box using the Softmax function.

[0073] In a possible implementation, the category prediction probability threshold used in step S103 is a pre-set category prediction probability threshold. In step S103, the category prediction probability of each second candidate box is compared with the same category prediction probability threshold. For a second candidate box, when the category prediction probability of the second candidate box is greater than the category prediction probability threshold, the second candidate box is determined to be the third candidate box; when the category prediction probability of the second candidate box is not greater than the category prediction probability threshold, the second candidate box is determined not to be the third candidate box.

[0074] For example, step S102 determines M second candidate boxes, including the second candidate box 1, the second candidate box 2, ... the second candidate box M. The category prediction probability threshold used in step S103 is recorded as the category prediction probability threshold A. The second candidate box 1 is compared with the category prediction probability threshold A, the second candidate box 2 is compared with the category prediction probability threshold A, and the second candidate box M is compared with the category prediction probability threshold A. When the category prediction probability of the second candidate box 1 is greater than the category prediction probability threshold A, the second candidate box 1 is determined as the third candidate box. When the category prediction probability of the second candidate box 2 is greater than the category prediction probability threshold A, the second candidate box 2 is determined as the third candidate box. When the category prediction probability of the second candidate box M is greater than the category prediction probability threshold A, the second candidate box M is determined as the third candidate box.

[0075] In step S104, the target detection result corresponding to the target image is determined according to each third candidate box.

[0076] In step S104, an algorithm for determining a detection frame for generating a target detection result corresponding to the target image, such as an NMS algorithm, may be used to determine a detection frame for generating a target detection result corresponding to the target image according to each third candidate frame. The detection frame for generating a target detection result corresponding to the target image is a two-dimensional frame.

[0077] As an example, in step S104, all third candidate boxes are deduplicated using the NMS algorithm to determine a detection box for generating a target detection result corresponding to the target image.

[0078] In a possible implementation, the target detection result corresponding to the target image is a corresponding two-dimensional detection result. The target detection result corresponding to the target image includes: a detection frame used to generate the target detection result corresponding to the target image, and a target category of the detection frame used to generate the target detection result corresponding to the target image.

[0079] Among them, the detection frame used to generate the target detection result corresponding to the target image can be represented by the coordinates of the center point of the detection frame used to generate the target detection result corresponding to the target image in the image coordinate system of the target image, the length of the detection frame used to generate the target detection result corresponding to the target image, and the width of the detection frame used to generate the target detection result corresponding to the target image.

[0080] In a possible implementation, the target detection result corresponding to the target image is a three-dimensional detection result. The target detection result corresponding to the target image includes: a three-dimensional frame corresponding to the detection frame used to generate the target detection result corresponding to the target image, the category of the target in the third candidate frame, and the orientation angle of the target in the third candidate frame. Among them, the three-dimensional frame corresponding to the detection frame used to generate the target detection result corresponding to the target image can be represented by the coordinates of the three-dimensional frame corresponding to the detection frame used to generate the target detection result corresponding to the target image in the world coordinate system. The three-dimensional frame corresponding to the detection frame used to generate the target detection result corresponding to the target image and the orientation angle of the target in the three-dimensional frame are determined by using the internal parameters and external parameters of the camera of the vehicle that collects the target image, the length of the detection frame used to generate the target detection result corresponding to the target image, the width of the detection frame used to generate the target detection result corresponding to the target image, the target category corresponding to the detection frame used to generate the target detection result corresponding to the target image, the coordinates of the center point of the detection frame used to generate the target detection result corresponding to the target image in the image coordinate system of the target image, and the target depth corresponding to the detection frame used to generate the target detection result corresponding to the target image.

[0081] refer to Figure 2 , which shows a flowchart of an example of post-processing of target detection using the target detection method provided by an embodiment of the present application.

[0082] In this example, two stages of candidate box screening are performed. In the first stage of candidate box screening, M1 second candidate boxes are screened from N1 first candidate boxes. Among them, M1 is less than N1. In the second stage of candidate box screening, L1 third candidate boxes are screened from M1 second candidate boxes based on the category prediction probabilities and category prediction probability thresholds of the M1 second candidate boxes. In the second stage of candidate box screening, the category prediction probability of each second candidate box is compared with the same category prediction probability threshold. In this example, the NMS algorithm is used to determine K1 detection boxes for determining the target detection result based on each third candidate box in the L1 third candidate boxes.

[0083] refer to Figure 3 , which shows a flow chart of another target detection method provided in an embodiment of the present application.

[0084] In step S301, the target image is input into the target detection network, and the target detection network outputs prediction score information of multiple first candidate boxes.

[0085] In step S302, for each first candidate box, when the existence prediction score of the first candidate box is greater than the existence prediction score threshold and the target category prediction score of the first candidate box is greater than the category prediction score threshold, the first candidate box is used as the second candidate box.

[0086] In step S302, for each first candidate box, when the existence prediction score of the first candidate box is not greater than the existence prediction score threshold and / or the target category prediction score of the first candidate box is not greater than the category prediction score threshold, it can be determined that the first candidate box is not used as the second candidate box.

[0087] For example, for the first candidate box A, when the existence prediction score of the first candidate box A is greater than the existence prediction score threshold and the target category prediction score of the first candidate box A is greater than the category prediction score threshold, the first candidate box A is used as the second candidate box; when the existence prediction score of the first candidate box A is not greater than the existence prediction score threshold and / or the target category prediction score of the first candidate box A is not greater than the category prediction score threshold, it is determined that the first candidate box is not used as the second candidate box.

[0088] In step S303, for each second candidate box, it is determined whether the second candidate box is used as the third candidate box according to the category prediction probability of the second candidate box and the category prediction probability threshold corresponding to the second candidate box.

[0089] Step S303 includes: step S3031-step S3033.

[0090] In step S3031, for each second candidate box, the target category prediction score of the second candidate box is converted into a category prediction probability of the second candidate box.

[0091] In step S3032, for each second candidate box, a category prediction probability threshold corresponding to the second candidate box is determined according to the associated information of the second candidate box.

[0092] For example, step S302 determines M second candidate boxes, including second candidate box 1, second candidate box 2, ... second candidate box M. In step S3032, based on the association information of second candidate box 1, the category prediction probability threshold corresponding to second candidate box 1 is determined. Based on the association information of second candidate box 2, the category prediction probability threshold corresponding to second candidate box 2 is determined. Based on the association information of second candidate box M, the category prediction probability threshold corresponding to second candidate box M is determined.

[0093] For a second candidate box, the associated information of the second candidate box includes at least some of the following items: a target depth corresponding to the second candidate box, a target category corresponding to the second candidate box, and a scale of a target feature map corresponding to the second candidate box.

[0094] It should be noted that, when the target detection network is a multi-scale prediction model, the associated information of the second candidate box may include the scale of the target feature map corresponding to the second candidate box.

[0095] In the embodiment of the present application, the target depth corresponding to the second candidate frame may refer to: the depth corresponding to the target at least partially surrounded by the second candidate frame.

[0096] The target depth corresponding to the second candidate frame may be: the depth corresponding to the center point of the second candidate frame. The target depth corresponding to the second candidate frame may be predicted by the target detection network.

[0097] The target depth corresponding to the second candidate frame may indicate: the distance between a camera that captures the target image at the moment when the target image is captured and an object in the real world represented by at least a portion of the target surrounded by the second candidate frame.

[0098] In an embodiment of the present application, the target category corresponding to the second candidate box is: the category for which the target category prediction score of the second candidate box is targeted among multiple categories.

[0099] As an example, the multiple categories consist of category 1, category 2...category N. For a first candidate box, the prediction score information of the first candidate box includes: the category prediction score of the first candidate box for category 1, the category prediction score of the first candidate box for category 2...the category prediction score of the first candidate box for category N. Among them, for category j among the multiple categories, the category prediction score of the first candidate box for category j is greater than the score of the first candidate box for any other category among the multiple categories except category j, and the category prediction score of the first candidate box for category j is: the target category prediction score of the first candidate box. The target category corresponding to the second candidate box is: category j.

[0100] For a second candidate box, the target feature map corresponding to the second candidate box is: a feature map used when predicting a target category prediction score of the second candidate box.

[0101] The feature map used when predicting the target category prediction score of the second candidate box can be: a feature map used as an input to a detection head of a target detection network when predicting the target category prediction score of the second candidate box. When predicting the target category prediction score of the second candidate box, the feature map used when predicting the target category prediction score of the second candidate box can be input to a detection head of the target detection network corresponding to the scale of the target feature map corresponding to the second candidate box, and the detection head outputs the target category prediction score of the second candidate box.

[0102] Step S3032 takes into account that the target category corresponding to the candidate box, the target depth corresponding to the candidate box, and the scale of the target feature map corresponding to the candidate box are all associated with the category prediction probability of the candidate box. When determining the category prediction probability threshold corresponding to the second candidate box, the target category corresponding to the second candidate box, the target depth corresponding to the second candidate box, and the scale of the target feature map corresponding to the second candidate box are used to determine the category prediction probability threshold corresponding to the second candidate box, and the category prediction probability threshold corresponding to the second candidate box is determined to be suitable for comparison with the category prediction probability of the second candidate box. Thus, it is possible to accurately determine whether the second candidate box is used as the third candidate box, improve the accuracy of the determined third candidate box, and improve the accuracy of the target detection result. The determined category prediction probability threshold corresponding to the second candidate box can avoid the situation where the third candidate box is inaccurate due to the use of a fixed category prediction probability threshold for comparison of the category prediction probability of each second candidate box, thereby improving the accuracy of the target detection result. Among them, the inaccurate situation of the determined third candidate box includes: the second candidate box that can be used as the third candidate box is missed, that is, it is not determined as the third candidate box.

[0103] It should be noted that the second candidate box is usually missed because the target category corresponding to the second candidate box is a rare category and / or the target depth corresponding to the candidate box is deep and / or the size of at least a part of the target surrounded by the second candidate box is small.

[0104] As an example, the correlation between the target category corresponding to the candidate box and the category prediction probability of the candidate box can be: when the target depth corresponding to the candidate box and the scale of the target feature map corresponding to the candidate box are both constants, the lower the proportion of the target category corresponding to the candidate box in the training set of the target detection network, the lower the category prediction probability of the candidate box.

[0105] Among them, the proportion of the target category corresponding to the candidate box in the training set of the target detection network can be: the number of labeled targets of the target category corresponding to the candidate box in the training set of the target detection network divided by the total number of labeled targets in the training set of the target detection network.

[0106] It should be noted that when the proportion of the target category corresponding to the candidate box in the training set of the target detection network is less than the proportion threshold, the target category corresponding to the candidate box can be called a rare category.

[0107] As an example, the correlation between the target depth corresponding to the candidate box and the category prediction probability of the candidate box can be: when the target category corresponding to the candidate box and the scale of the target feature map corresponding to the candidate box are constants, the deeper the target depth corresponding to the candidate box, the lower the category prediction probability of the candidate box.

[0108] As an example, the correlation relationship between the scale of the target feature map corresponding to the candidate box and the category prediction probability of the candidate box can be: when the target depth corresponding to the candidate box and the target category corresponding to the candidate box are both constants, the smaller the scale of the target feature map corresponding to the candidate box, the lower the category prediction probability of the candidate box.

[0109] In an embodiment of the present application, in order to determine the category prediction probability threshold corresponding to the second candidate box, the association probability of each category in the multiple categories can be preset. In one possible way, for a category, if the proportion of the category in the training set of the target detection network is less than the proportion threshold, the association probability of the category is a pre-set probability greater than 0 and less than 1; if the proportion of the category in the training set of the target detection network is not less than the proportion threshold, the association probability of the category can be 0. Multiple depth intervals can be preset, each depth interval has an association probability, and the right endpoint of the i-th depth interval is the left endpoint of the i+1-th depth interval. The left endpoint of the i-th depth interval is the right endpoint of the i-1-th depth interval. The depth intervals other than the depth interval with the largest left endpoint can be a left-open and right-closed interval. The depth interval with the largest left endpoint can be a left-open and right-open interval, and the right endpoint of the depth interval with the largest left endpoint can be positive infinity. In a possible implementation, the larger the left endpoint value of the depth interval, the greater the association probability of the depth interval. As an example, the association probability of the depth interval with the smallest left endpoint can be 0. The association probability of each feature map scale in the plurality of feature map scales may be preset. In a possible implementation, the larger the feature map scale, the smaller the association probability of the feature map scale. As an example, the association probability of the largest feature map scale may be 0.

[0110] In an embodiment of the present application, in order to determine the category prediction probability threshold corresponding to the second candidate box, a basic category prediction probability threshold may be pre-set.

[0111] In a possible implementation of step S3032, step S3032 includes: step S30321. In step S30321, the probability of each information item in the associated information of the second candidate box being used to determine the category prediction probability threshold corresponding to the second candidate box is determined.

[0112] Among them, the probability of the target depth corresponding to the second candidate box for determining the category prediction probability threshold corresponding to the second candidate box is: the opposite number of the associated probability of the target depth corresponding to the second candidate box. The probability of the target category corresponding to the second candidate box for determining the category prediction probability threshold corresponding to the second candidate box is: the associated probability of the target category corresponding to the second candidate box. The probability of the scale of the target feature map corresponding to the second candidate box for determining the category prediction probability threshold corresponding to the second candidate box is: the opposite number of the associated probability of the scale of the target feature map corresponding to the second candidate box.

[0113] The association probability of the target depth corresponding to the second candidate frame is: the association probability of the depth interval in which the target depth corresponding to the second candidate frame is located.

[0114] In step S30321, the sum of the probabilities of each information item in the associated information of the second candidate box for determining the category prediction probability threshold corresponding to the second candidate box is determined, and the basic category prediction probability threshold is added to the sum to obtain the category prediction probability threshold corresponding to the second candidate box.

[0115] The method of obtaining the category prediction probability threshold corresponding to the second candidate box in step S30321 can be expressed as: the category prediction probability threshold corresponding to the second candidate box = the basic category prediction probability threshold - the association probability of the scale of the target feature map corresponding to the second candidate box - the association probability of the target depth corresponding to the second candidate box + the association probability of the target category corresponding to the second candidate box.

[0116] In another possible implementation of step S3032, step S3032 includes: step S30322. In step S30322, for a second candidate box, when the target category corresponding to the second candidate box is a vehicle category, the category prediction probability threshold corresponding to the second candidate box is determined according to the scale of the target feature map corresponding to the second candidate box; when the target category corresponding to the second candidate box is an obstacle category or a vulnerable road user (Vulnerable Road Users, VRU) category, the category prediction probability threshold corresponding to the second candidate box is determined according to the target depth corresponding to the second candidate box.

[0117] Step S30322 takes into account that: the scale of the target feature map corresponding to the second candidate box has a high correlation with the category prediction probability of the second candidate box surrounding at least a portion of the vehicle. The target depth corresponding to the second candidate box has a high correlation with the category prediction probability of the second candidate box surrounding at least a portion of the obstacle or vulnerable road user. Step S30322 determines the category prediction probability threshold corresponding to the second candidate box using information with a high correlation with the category prediction probability of the second candidate box, thereby improving the accuracy of determining the category prediction probability threshold corresponding to the second candidate box.

[0118] In a possible implementation of step S30322, for a second candidate frame, when the target category corresponding to the second candidate frame is a vehicle category, the association probability of the scale of the target feature map corresponding to the second candidate frame is determined, and the first difference obtained by subtracting the association probability of the scale of the target feature map corresponding to the second candidate frame from the basic category prediction probability threshold is determined as the category prediction probability threshold corresponding to the second candidate frame. When the target category corresponding to the second candidate frame is an obstacle category or a vulnerable road user category, the second difference obtained by subtracting the association probability of the target depth corresponding to the second candidate frame from the basic category prediction probability threshold is determined, and the sum of the second difference and the association probability of the target category corresponding to the second candidate frame is determined, and the sum is determined as the category prediction probability threshold corresponding to the second candidate frame. Among them, the association probability of the target depth corresponding to the second candidate frame is: the association probability of the depth interval in which the target depth corresponding to the second candidate frame is located.

[0119] In another possible implementation of step S30312, for a second candidate box, when the target category corresponding to the second candidate box is a vehicle category, the category prediction probability threshold corresponding to the second candidate box is determined according to the scale of the target feature map corresponding to the second candidate box and the feature map scale weighting coefficient. When the target category corresponding to the second candidate box is an obstacle category or a vulnerable road user category, the category prediction probability threshold corresponding to the second candidate box is determined according to the target depth corresponding to the second candidate box, the depth weighting coefficient, and the category weighting coefficient.

[0120] The feature map scale weighting coefficient, depth weighting coefficient, and category weighting coefficient are all preset. As an example, the feature map scale weighting coefficient, depth weighting coefficient, and category weighting coefficient are all in the range of 0.1 to 0.3.

[0121] As an example, in another possible implementation of step S30312, a method for determining a category prediction probability threshold corresponding to a second candidate box can be expressed as:

[0122]

[0123] Among them, Thr baseis the prediction probability threshold of the basic category, w depth is the depth weighting coefficient, w cls is the category weighting coefficient, w level is the feature map scale weighting coefficient, D is the associated value of the target depth corresponding to the second candidate box, and L is the associated value of the scale of the target feature map corresponding to the second candidate box;

[0124] if cls∈VRU or Obstracle, it means that the target category corresponding to the second candidate box is an obstacle category or a vulnerable road user category;

[0125] If cls∈Vehicle, it means that the target category corresponding to the second candidate box is the vehicle category.

[0126] As an example, the associated value of the target depth corresponding to the second candidate frame is in the range of 0 to 3. The associated value of the scale of the target feature map corresponding to the second candidate frame is in the range of 0 to 3.

[0127] For a second candidate frame, the associated value of the target depth corresponding to the second candidate frame is: the associated value of the depth interval in which the target depth corresponding to the second candidate frame is located.

[0128] Each depth interval has an associated value. In a possible implementation, the larger the left endpoint value of the depth interval, the larger the associated value of the depth interval. As an example, the associated value of the depth interval with the smallest left endpoint can be 0. The associated value of each feature map scale in the multiple feature map scales can be preset. In a possible implementation, the larger the feature map scale, the smaller the associated value of the feature map scale. As an example, the associated value of the largest feature map scale can be 0.

[0129] In step S3033, for each second candidate box, it is determined whether the second candidate box is used as the third candidate box according to the category prediction probability of the second candidate box and the category prediction probability threshold corresponding to the second candidate box.

[0130] In one possible implementation, in step S3033, for a second candidate box, when the category prediction probability of the second candidate box is greater than the category prediction probability threshold corresponding to the second candidate box, the second candidate box is determined to be the third candidate box; when the category prediction probability of the second candidate box is not greater than the category prediction probability threshold corresponding to the second candidate box, the second candidate box is determined not to be the third candidate box.

[0131] For example, step S302 is used to determine M second candidate boxes, including the second candidate box 1, the second candidate box 2, ... the second candidate box M. In step S3033, the second candidate box 1 is compared with the category prediction probability threshold corresponding to the second candidate box 1, the second candidate box 2 is compared with the category prediction probability threshold corresponding to the second candidate box 1, and the second candidate box M is compared with the category prediction probability threshold corresponding to the second candidate box 1. When the second candidate box 1 is greater than the category prediction probability threshold corresponding to the second candidate box 1, the second candidate box 1 is determined as the third candidate box. When the second candidate box 2 is greater than the category prediction probability threshold corresponding to the second candidate box 2, the second candidate box 2 is determined as the third candidate box. When the second candidate box M is greater than the category prediction probability threshold corresponding to the second candidate box M, the second candidate box M is determined as the third candidate box.

[0132] In step S304, the target detection result corresponding to the target image is determined according to each third candidate box.

[0133] The process of step S304 refers to the process of step S104.

[0134] refer to Figure 4 , which shows a flowchart of another example of post-processing of target detection by the target detection method provided by an embodiment of the present application.

[0135] In this example, two stages of candidate box screening are performed. In the first stage of candidate box screening, M2 second candidate boxes are screened from N2 first candidate boxes. Among them, M2 is less than N2. In the second stage of candidate box screening, L2 third candidate boxes are screened from M2 second candidate boxes based on the category prediction probabilities of the M2 second candidate boxes and the category prediction probability threshold corresponding to each second candidate box in the M2 second candidate boxes. In the second stage of candidate box screening, the category prediction probability of each second candidate box is compared with the category prediction probability threshold corresponding to each second candidate box. In this example, the NMS algorithm is used to determine K2 detection boxes for determining the target detection result based on each third candidate box in the L2 third candidate boxes.

[0136] The embodiments of the present application provide a target detection device. The device is used to implement the above embodiments and preferred embodiments, and the descriptions that have been made will not be repeated. As used below, the term "unit" can implement a combination of software and / or hardware of a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceivable.

[0137] The target detection device is installed on a vehicle, and the target detection device comprises:

[0138] An input unit, used to input the target image into the target detection network to obtain prediction score information of multiple first candidate boxes, wherein the prediction score information of the first candidate boxes includes: a category prediction score and an existence prediction score for each category in the multiple categories;

[0139] A first screening unit is used to screen out a second candidate frame from multiple first candidate frames according to a target category prediction score and an existence prediction score of each first candidate frame, a category prediction score threshold, and an existence prediction score threshold, wherein the target category prediction score of the first candidate frame is the largest score among the multiple category prediction scores of the first candidate frames;

[0140] A second screening unit is used to convert the target category prediction score of the second candidate box into a category prediction probability of the second candidate box for each second candidate box, and determine whether the second candidate box is used as the third candidate box according to the category prediction probability of the second candidate box and the category prediction probability threshold;

[0141] The target detection result determination unit is used to determine the target detection result corresponding to the target image according to each third candidate frame.

[0142] In one possible implementation, the first screening unit is further used to, for each first candidate box, use the first candidate box as the second candidate box when the existence prediction score of the first candidate box is greater than the existence prediction score threshold and the target category prediction score of the first candidate box is greater than the category prediction score threshold.

[0143] In a possible implementation, the second screening unit is also used to determine a category prediction probability threshold corresponding to the second candidate box based on the association information of the second candidate box, the association information including at least some of the following items: a target depth corresponding to the second candidate box, a target category corresponding to the second candidate box, and a scale of a target feature map corresponding to the second candidate box, the target category is a category for which the target category prediction score of the second candidate box is targeted, and the target feature map is a feature map used when predicting the target category prediction score of the second candidate box; determine whether the second candidate box is used as the third candidate box based on the category prediction probability of the second candidate box and the category prediction probability threshold corresponding to the second candidate box.

[0144] In a possible implementation, the second screening unit is also used to determine the category prediction probability threshold corresponding to the second candidate box according to the scale of the target feature map corresponding to the second candidate box when the target category corresponding to the second candidate box is a vehicle category; when the target category corresponding to the second candidate box is an obstacle category or a vulnerable road user category, determine the category prediction probability threshold corresponding to the second candidate box according to the target depth corresponding to the second candidate box.

[0145] In a possible implementation, the second screening unit is also used to determine the category prediction probability threshold corresponding to the second candidate box according to the scale of the target feature map corresponding to the second candidate box and the feature map scale weighting coefficient when the target category corresponding to the second candidate box is a vehicle category.

[0146] In one possible implementation, the second screening unit is also used to determine the category prediction probability threshold corresponding to the second candidate box based on the target depth, depth weighting coefficient, and category weighting coefficient corresponding to the second candidate box when the target category corresponding to the second candidate box is an obstacle category or a vulnerable road user category.

[0147] In this embodiment, the device is presented in the form of a functional unit, where the unit refers to an ASIC circuit, a processor and a memory that executes one or more software or fixed programs, and / or other devices that can provide the above functions.

[0148] The further functional description of each of the above units is the same as that of the above corresponding embodiments and will not be repeated here.

[0149] refer to Figure 5 , which shows a schematic diagram of the structure of a computer device provided by an embodiment of the present application, the computer device is installed on a vehicle, and the computer device includes: one or more processors 10, a memory 20, and an interface for connecting various components, including a high-speed interface and a low-speed interface. The various components are connected to each other using different buses for communication, and can be installed on a common mainboard or installed in other ways as needed. The processor can process instructions executed in the computer device, including instructions stored in or on the memory to display graphical information of the GUI on an external input / output device (such as a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Similarly, multiple computer devices can be connected, and each device provides some necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system).

[0150] The processor 10 may be a central processing unit, a network processor or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be a dedicated integrated circuit, a programmable logic device or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic or any combination thereof.

[0151] The memory 20 stores instructions executable by at least one processor 10, so that the at least one processor 10 executes the method shown in the above embodiment.

[0152] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely arranged relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0153] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid state drive; the memory 20 may also include a combination of the above types of memory.

[0154] The computer device further includes an input device 30 and an output device 40. The processor 10, the memory 20, the input device 30 and the output device 40 may be connected via a bus or other means.

[0155] The input device 30 can receive input digital or character information, and generate key signal input related to the user settings and function control of the computer device, such as a touch screen, a keypad, a mouse, a track pad, a touch pad, an indicator bar, one or more mouse buttons, a trackball, a joystick, etc. The output device 40 may include a display device, an auxiliary lighting device (e.g., an LED) and a tactile feedback device (e.g., a vibration motor), etc. The above-mentioned display device includes but is not limited to a liquid crystal display, a light emitting diode, a display and a plasma display. In some optional embodiments, the display device can be a touch screen.

[0156] The embodiment of the present application also provides a computer-readable storage medium. The above method according to the embodiment of the present application can be implemented in hardware, firmware, or can be implemented as a computer code that can be recorded in a storage medium, or can be implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and will be stored in a local storage medium through a network download, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state hard disk, etc.; further, the storage medium can also include a combination of the above types of memories. It can be understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor, or hardware, the method shown in the above embodiment is implemented.

[0157] A part of the embodiments of the present application may be applied as a computer program product, such as a computer program instruction, which, when executed by a computer, can be called or provided through the operation of the computer. It should be understood by those skilled in the art that the existence of computer program instructions in a computer-readable medium includes, but is not limited to, source files, executable files, installation package files, etc., and accordingly, the way in which the computer program instructions are executed by the computer includes, but is not limited to: the computer directly executes the instruction, or the computer compiles the instruction and then executes the corresponding compiled program, or the computer reads and executes the instruction, or the computer reads and installs the instruction and then executes the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to the computer.

[0158] Although the embodiments of the present application are described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present application, and such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A target detection method, characterized in that: The method comprises: Inputting the target image into the target detection network to obtain prediction score information of multiple first candidate boxes, the prediction score information of the first candidate boxes including: a category prediction score and an existence prediction score for each category in the multiple categories; According to the target category prediction score and existence prediction score of each first candidate box, the category prediction score threshold, and the existence prediction score threshold, a second candidate box is screened out from the multiple first candidate boxes, wherein the target category prediction score of the first candidate box is the largest score among the multiple category prediction scores of the first candidate boxes; For each second candidate box, converting the target category prediction score of the second candidate box into the category prediction probability of the second candidate box, and determining whether the second candidate box is used as the third candidate box according to the category prediction probability of the second candidate box and the category prediction probability threshold; According to each third candidate box, a target detection result corresponding to the target image is determined.

2. The method according to claim 1, characterized in that: According to the target category prediction score and the existence prediction score of each first candidate box, the category prediction score threshold, and the existence prediction score threshold, screening out the second candidate box from the multiple first candidate boxes includes: For each first candidate box, when the existence prediction score of the first candidate box is greater than the existence prediction score threshold and the target category prediction score of the first candidate box is greater than the category prediction score threshold, the first candidate box is used as the second candidate box.

3. The method according to claim 1, characterized in that According to the category prediction probability and the category prediction probability threshold of the second candidate box, determining whether the second candidate box is included as the third candidate box: Determine, according to the association information of the second candidate box, a category prediction probability threshold corresponding to the second candidate box, the association information including at least some of the following items: a target depth corresponding to the second candidate box, a target category corresponding to the second candidate box, and a scale of a target feature map corresponding to the second candidate box, the target category is a category for which the target category prediction score of the second candidate box is targeted, and the target feature map is a feature map used when predicting the target category prediction score of the second candidate box; According to the category prediction probability of the second candidate box and the category prediction probability threshold corresponding to the second candidate box, it is determined whether the second candidate box is used as the third candidate box.

4. The method according to claim 3, characterized in that: Determining, according to the association information of the second candidate box, a category prediction probability threshold corresponding to the second candidate box includes: When the target category corresponding to the second candidate box is a vehicle category, determining a category prediction probability threshold corresponding to the second candidate box according to the scale of the target feature map corresponding to the second candidate box; When the target category corresponding to the second candidate box is an obstacle category or a vulnerable road user category, a category prediction probability threshold corresponding to the second candidate box is determined according to the target depth corresponding to the second candidate box.

5. The method according to claim 4, characterized in that According to the scale of the target feature map corresponding to the second candidate box, determining the category prediction probability threshold corresponding to the second candidate box includes: According to the scale of the target feature map corresponding to the second candidate box and the feature map scale weighting coefficient, the category prediction probability threshold corresponding to the second candidate box is determined.

6. The method according to claim 4, characterized in that Determining the category prediction probability threshold corresponding to the second candidate box according to the target depth corresponding to the second candidate box includes: According to the target depth, the depth weighting coefficient, and the category weighting coefficient corresponding to the second candidate box, a category prediction probability threshold corresponding to the second candidate box is determined.

7. A target detection device, characterized in that: Installed on a vehicle, the device comprises: An input unit, used to input the target image into the target detection network to obtain prediction score information of multiple first candidate boxes, wherein the prediction score information of the first candidate boxes includes: a category prediction score and an existence prediction score for each category in the multiple categories; A first screening unit is used to screen out a second candidate frame from multiple first candidate frames according to a target category prediction score and an existence prediction score of each first candidate frame, a category prediction score threshold, and an existence prediction score threshold, wherein the target category prediction score of the first candidate frame is the largest score among the multiple category prediction scores of the first candidate frames; A second screening unit is used to convert the target category prediction score of the second candidate box into a category prediction probability of the second candidate box for each second candidate box, and determine whether the second candidate box is used as the third candidate box according to the category prediction probability of the second candidate box and the category prediction probability threshold; The target detection result determination unit is used to determine the target detection result corresponding to the target image according to each third candidate frame.

8. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the method according to any one of claims 1 to 6 by executing the computer instructions.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the method according to any one of claims 1 to 6.

10. A computer program product, characterized in that The method comprises computer instructions for causing a computer to execute the method according to any one of claims 1 to 6.

Images