A target detection method, a computer readable storage medium and a driving device

Abstract

The present application relates to the technical field of computer, and specifically provides a target detection method, a computer readable storage medium and a driving device, in the training process of a target detection model, a preliminary identification category of an anchor box is obtained according to the anchor box and a label box of a training image sample; for the anchor box with the preliminary identification category as a positive sample, the final identification category of the anchor box with the preliminary identification category as the positive sample is further determined according to the number of label boxes with an intersection-over-union greater than a preset intersection-over-union threshold, or based on the number of label boxes with an intersection-over-union greater than the preset intersection-over-union threshold and the position relationship between the anchor box and the label box; and the target detection model to be trained is trained at least according to the anchor box, the final identification category of the anchor box and the label box, to obtain a trained target detection model. Based on the trained target detection model, target detection can be performed, which can effectively improve the detection effect and avoid the problem of overlapping boxes.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically providing a target detection method, a computer-readable storage medium, and a driving device. Background Technology

[0002] Object detection is a fundamental and crucial module in intelligent vehicle driver assistance systems. The goal of object detection is to detect as many obstacles of interest as possible, ensuring no overlap or omission. However, in many scenarios, especially those with dense obstacles, existing detection methods are prone to generating false detection boxes between two obstacles, known as duplicate boxes. The main reason for this phenomenon is that current detection models often assign positive and negative samples during the training phase based on the intersection-union (IoU) ratio between the anchor box and the labeled box, or whether the current point is within the labeled box. These conditions are also met for boxes that span two objects or enclose two objects, and there are indeed some obvious obstacle features within these boxes, making duplicate boxes a likely result. Summary of the Invention

[0003] The present invention aims to solve the above-mentioned technical problem, namely, to solve the problem that existing target detection methods are prone to overlapping bounding boxes in scenes with dense targets.

[0004] In a first aspect, the present invention provides a target detection method, comprising:

[0005] Acquire the image to be detected;

[0006] The image to be detected is input into the trained target detection model to obtain the target detection result;

[0007] The object detection model is trained based on the following steps:

[0008] Obtain the bounding boxes and anchor boxes of the training image samples, and determine the preliminary recognition category of the anchor boxes based on the bounding boxes;

[0009] For the anchor box whose initial identification category is positive, the final identification category of the anchor box whose initial identification category is positive is determined based on the number of bounding boxes whose intersection-union ratio with the anchor box is greater than a preset intersection-union ratio threshold, or based on the number of bounding boxes whose intersection-union ratio with the anchor box is greater than a preset intersection-union ratio threshold and the positional relationship between the anchor box and the bounding boxes, wherein the final identification category includes target positive sample or target negative sample;

[0010] The target detection model is trained based at least on the anchor boxes, the final recognition category of the anchor boxes, and the labeled boxes to obtain a trained target detection model.

[0011] In some embodiments, determining the final identification category of the anchor boxes whose initial identification category is positive based on the number of bounding boxes whose intersection-union ratio with the anchor box is greater than a preset intersection-union ratio threshold includes:

[0012] When the number of bounding boxes whose intersection-union ratio with the anchor box is greater than a preset intersection-union ratio threshold is greater than 1, the final identification category of the anchor box is determined to be the target negative sample.

[0013] In some embodiments, determining the final identification category of the anchor boxes whose initial identification category is positive based on the number of bounding boxes whose intersection-union ratio (IU) with the anchor boxes is greater than a preset IU threshold and the positional relationship between the anchor boxes and the bounding boxes includes:

[0014] The final identification category of the anchor box whose initial identification category is positive is determined based on the number of bounding boxes whose intersection-union ratio with the anchor box is greater than a preset intersection-union ratio threshold, and the positional relationship between the center of the anchor box and the preset center region of the bounding box.

[0015] In some embodiments, determining the final identification category of the anchor boxes whose initial identification category is positive based on the number of bounding boxes whose intersection-union ratio (IU) with the anchor boxes is greater than a preset IU threshold, and the positional relationship between the center of the anchor boxes and a preset center region of the bounding boxes, includes:

[0016] When the number of annotation boxes with an intersection-union ratio greater than a preset intersection-union ratio threshold is 1, it is determined whether the center of the anchor box is in the preset center region of the annotation box.

[0017] When the center of the anchor box is within the preset center area of ​​the annotation box, the final identification category of the anchor box is determined to be a target positive sample; when the center of the anchor box is not within the preset center area of ​​the annotation box, the final identification category of the anchor box is determined to be a target negative sample.

[0018] In some embodiments, training the target detection model based at least on the anchor boxes, the final identified categories of the anchor boxes, and the labeled boxes includes:

[0019] Based on the anchor boxes that are ultimately identified as positive target samples and / or negative target samples, calculate the loss function between the anchor boxes and the labeled boxes, and train the target detection model to be trained according to the loss function;

[0020] or,

[0021] Based on the anchor boxes whose final identification category is positive target sample and / or negative target sample, and the anchor boxes whose initial identification category is negative sample, a loss function is calculated between the anchor boxes and the labeled boxes, so as to train the target detection model to be trained according to the loss function.

[0022] In some embodiments, the preset central region is constructed through the following steps:

[0023] Obtain the border length and border width of the annotation box;

[0024] With the center point of the annotation box as the center, and the border length at a first ratio as the side length and the border width at a second ratio as the side width, a rectangular preset central region is constructed, wherein the first ratio and the second ratio are less than 1.

[0025] In some embodiments, the preset central region is constructed through the following steps:

[0026] A pre-defined circular central region is constructed with the center point of the annotation box as the center and the third ratio of the minimum side length of the annotation box as the radius, wherein the third ratio is less than 1.

[0027] In some embodiments, the preset central region is constructed through the following steps:

[0028] With the center point of the annotation box as the center, and the fourth ratio of the longest side length of the annotation box as the first axis and the fifth ratio of the shortest side length of the annotation box as the second axis, an elliptical preset central region is constructed, wherein the fourth ratio and the fifth ratio are less than 1.

[0029] In a second aspect, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the target detection method described in any one of the preceding claims.

[0030] In a third aspect, the present invention provides a driving device, which includes a driving device body, a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements the target detection method described in any one of the above.

[0031] By employing the above technical solution, during the training process of the target detection model, this invention obtains the preliminary identification category of the anchor boxes based on the labeled boxes and anchor boxes of the training image samples. For anchor boxes whose preliminary identification category is positive, the final identification category of the anchor boxes whose preliminary identification category is positive is further determined based on the number of labeled boxes whose intersection-union ratio (IU) with the anchor boxes is greater than a preset IU threshold, or based on the number of labeled boxes whose IU with the anchor boxes is greater than the preset IU threshold and the positional relationship between the anchor boxes and the labeled boxes. At least based on the anchor boxes, the final identification category of the anchor boxes, and the labeled boxes, the target detection model to be trained is trained to obtain a trained target detection model. This invention, based on this trained target detection model, can effectively improve the detection performance and avoid the problem of duplicate bounding boxes. Attached Figure Description

[0032] The preferred embodiments of the present invention are described below with reference to the accompanying drawings, in which:

[0033] Figure 1 This is a schematic flowchart of a target detection method provided in an embodiment of the present invention;

[0034] Figure 2 This is a flowchart illustrating the target detection model training method provided in an embodiment of the present invention;

[0035] Figures 3 to 5 This is a schematic diagram of the preset center area of ​​the annotation box provided in a specific example of the present invention;

[0036] Figure 6 This is a schematic diagram of the detection bounding box obtained by pedestrian detection based on the relevant target detection method;

[0037] Figure 7 This is a schematic diagram of the detection frame obtained after pedestrian detection based on the target detection method provided in the embodiments of the present invention.

[0038] Figure 8 This is a schematic diagram of the training method for a target detection model provided in another embodiment of the present invention. Detailed Implementation

[0039] Some embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0040] Object detection is a fundamental and crucial module in intelligent vehicle driver assistance systems. The goal of object detection is to detect as many obstacles of interest as possible, ensuring no overlap or omission. However, in many scenarios, especially those with dense obstacles, existing detection methods are prone to generating false detection boxes between two obstacles, known as duplicate boxes. The main reason for this phenomenon is that current detection models often assign positive and negative samples during the training phase based on the intersection-union (IoU) ratio between the anchor box and the labeled box, or whether the current point is within the labeled box. These conditions are also met for boxes that span two objects or enclose two objects, and there are indeed some obvious obstacle features within these boxes, making duplicate boxes a likely result.

[0041] In view of this, the present invention provides an object detection method. During the training process of the object detection model, a preliminary identification category of the anchor boxes is obtained based on the bounding boxes and anchor boxes of the training image samples. For anchor boxes whose preliminary identification category is positive, the final identification category is determined based on the number of bounding boxes whose intersection-union ratio (IU) with the anchor box is greater than a preset IU threshold, or based on the number of bounding boxes whose IU with the anchor box is greater than the preset IU threshold and the positional relationship between the anchor box and the bounding boxes. The object detection model is trained at least based on the anchor boxes, the final identification category of the anchor boxes, and the bounding boxes to obtain a trained object detection model. The present invention, based on this trained object detection model, can effectively improve the detection performance and avoid the problem of duplicate bounding boxes.

[0042] See Figure 1 As shown, Figure 1 This is a flowchart illustrating a target detection method provided in an embodiment of the present invention, which may include:

[0043] Step S11: Acquire the image to be detected;

[0044] Step S12: Input the image to be detected into the trained target detection model to obtain the target detection result.

[0045] The target detection method provided in this invention can be applied to the detection of at least one target, such as vehicles, pedestrians, and animals.

[0046] In some embodiments, step S12 may specifically involve inputting the image to be detected into a trained target detection model to obtain a detection box used to indicate the target location, thus obtaining the target detection result.

[0047] See Figure 2 As shown, the object detection model is trained based on the following steps:

[0048] Step S21: Obtain the bounding boxes and anchor boxes of the training image samples and determine the preliminary recognition category of the anchor boxes based on the bounding boxes;

[0049] Step S22: For anchor boxes that are initially identified as positive samples, determine the final identification category of the anchor boxes that are initially identified as positive samples based on the number of annotation boxes whose intersection-union ratio with the anchor boxes is greater than a preset intersection-union ratio threshold, or based on the number of annotation boxes whose intersection-union ratio with the anchor boxes is greater than a preset intersection-union ratio threshold and the positional relationship between the anchor boxes and the annotation boxes. The final identification category includes target positive samples or target negative samples.

[0050] Step S23: Train the target detection model based at least on the anchor boxes, the final recognition category of the anchor boxes, and the labeled boxes to obtain a trained target detection model.

[0051] In some embodiments, step S21 may specifically be:

[0052] Obtain the bounding boxes and anchor points of the training image samples;

[0053] Calculate the intersection-union ratio (IUR) of the bounding box and the anchor box. When the IUR of the bounding box and the anchor box is greater than the preset IUR threshold, the category of the bounding box is determined as the category of the anchor box, and the anchor box is determined to be a positive sample or a negative sample relative to the category, that is, the preliminary identification category of the anchor box is determined.

[0054] In some embodiments, the final identification category of the anchor boxes whose initial identification category is positive based on the number of bounding boxes whose intersection-union ratio with the anchor boxes is greater than a preset intersection-union ratio threshold in step S22 can be specifically as follows:

[0055] When the number of bounding boxes whose intersection-union ratio with the anchor box is greater than the preset intersection-union ratio threshold is greater than 1, the final identification category of the anchor box is determined to be the target negative sample.

[0056] This method can filter anchor boxes that satisfy the cross-union ratio (CUI) greater than a preset threshold with multiple labeled boxes from anchor boxes initially identified as positive samples. The initial identification category of the anchor box is then corrected to determine the final identification category of the anchor box as a negative sample of the target. This removes some anchor boxes that are prone to causing duplicate boxes from the anchor boxes initially identified as positive samples, which is beneficial for efficient training of the target detection model and improving the detection accuracy of the trained target detection model. In turn, it can effectively avoid the problem of duplicate boxes when performing target detection based on the trained target detection model in the future.

[0057] In other embodiments, the final identification category of the anchor boxes whose initial identification category is positive can be determined in step S22 based on the number of bounding boxes whose intersection-union ratio with the anchor boxes is greater than a preset intersection-union ratio threshold and the positional relationship between the anchor boxes and the bounding boxes.

[0058] The final identification category of anchor boxes that are initially identified as positive samples is determined based on the number of bounding boxes whose intersection-union ratio (IU) with the anchor boxes is greater than a preset IU threshold, and the positional relationship between the center of the anchor box and the preset center region of the bounding box.

[0059] In other embodiments, the final identification category of an anchor box that is initially identified as a positive sample can be determined by combining whether the distance between the center of the anchor box and the center of the label box meets a preset distance threshold with the number of label boxes whose intersection-union ratio with the anchor box is greater than a preset intersection-union ratio threshold.

[0060] In other embodiments, the number of bounding boxes whose intersection-union ratio (IU) with the anchor box is greater than a preset IU threshold can be determined first. When the number of bounding boxes whose IU with the anchor box is greater than the preset IU threshold is 1, the final identification category of the anchor box whose initial identification category is positive is determined by combining the positional relationship between the center of the anchor box and the preset center region of the bounding box. For details, please refer to the following description of the present invention. Figure 8 The description.

[0061] In some embodiments, a preset central region can be constructed through the following steps:

[0062] Get the border length and border width of the annotation box;

[0063] Using the center point of the annotation box as the center, and the border length of the first ratio as the side length and the border width of the second ratio as the side width, construct a preset central area of ​​a rectangle, where the first ratio and the second ratio are less than 1.

[0064] As an example, when both the first and second scales are 0.5, and the border length of the annotation box is H and the border width is W, a rectangular preset center area with a length of 0.5H and a width of 0.5W is obtained based on the above construction method, such as... Figure 3 As shown.

[0065] In other embodiments, the preset central region can also be constructed through the following steps:

[0066] Construct a pre-defined circular central region with the center point of the annotation box as the center and the third ratio of the minimum side length of the annotation box as the radius, where the third ratio is less than 1.

[0067] As an example, when the third ratio is 0.5 and the border length H of the annotation box is greater than the border width W, a pre-defined circular central region is constructed with the center point of the annotation box as the center and a radius of 0.5W, based on the above construction method. Figure 4 As shown.

[0068] In other embodiments, the preset central region can also be constructed through the following steps:

[0069] Using the center point of the annotation box as the center, and the fourth ratio of the longest side of the annotation box as the first axis and the fifth ratio of the shortest side of the annotation box as the second axis, construct a preset central region of an ellipse, where the fourth and fifth ratios are less than 1.

[0070] As an example, when both the fourth and fifth scales are 0.5, and the border length H of the annotation box is greater than the border width W, based on the above construction method, an elliptical preset central region is obtained, centered on the center point of the annotation box, with 0.5H as the first axis and 0.5W as the second axis, as shown below. Figure 5 As shown.

[0071] In some embodiments, step S23 may specifically involve calculating the loss function between the anchor boxes and the labeled boxes based on the anchor boxes that are ultimately identified as positive target samples and / or negative target samples, so as to train the target detection model to be trained according to the loss function.

[0072] Using anchor boxes that identify positive targets as the final target category to train the target detection model can effectively improve the detection accuracy of the trained target detection model.

[0073] In addition, for anchor boxes that are initially identified as positive samples and are ultimately identified as negative samples, the target detection model to be trained can be trained based on these anchor boxes to obtain a target detection model with high detection accuracy. Using this target detection model for target detection can effectively avoid the problem of overlapping boxes.

[0074] It should be noted that for anchor boxes that are initially identified as positive samples and are ultimately identified as negative samples, these anchor boxes are prone to causing overlapping issues. When the overlap between these anchor boxes and anchor boxes labeled as positive samples is severe, in order to avoid interference during training, the anchor boxes ultimately identified as negative samples can be labeled as ignored samples, and the target detection model to be trained will not be based on anchor boxes labeled as ignored samples.

[0075] In other embodiments, step S23 may specifically involve calculating the loss function between the anchor boxes and the labeled boxes based on the anchor boxes that are ultimately identified as positive and / or negative samples of the target, and the anchor boxes that are initially identified as negative samples, so as to train the target detection model to be trained according to the loss function.

[0076] As an example, training the target detection model based on the anchor boxes and labeled boxes that are ultimately identified as positive target samples can be done as follows:

[0077] Obtain the offset of the annotation box relative to the anchor box whose final recognition category is the target positive sample, and annotate the anchor box according to the offset and the final recognition category;

[0078] Input the anchor box into the target detection model to be trained to obtain the predicted recognition category and predicted offset of the anchor box;

[0079] The class loss is calculated based on the predicted and final identified categories of the anchor boxes and the labeled categories, and the positional offset loss is calculated based on the predicted and offset values ​​of the anchor boxes and the labeled offset values. The target detection model is then trained based on the class loss and the positional offset loss until the model converges, resulting in a trained target detection model.

[0080] The trained object detection model can effectively improve the detection results when used to detect images. Existing object detection methods often suffer from overlapping bounding boxes in scenes with densely packed objects; the object detection method provided in this invention can effectively avoid this problem. See also [example description]. Figure 6 and Figure 7 As shown, Figure 6 To perform pedestrian detection using relevant object detection methods, three detection boxes were output for two pedestrians that were close together in the image, resulting in a problem of overlapping detection boxes. Figure 7 To perform pedestrian detection using the target detection method provided in this embodiment of the invention, detection boxes corresponding to two pedestrians are obtained respectively, thus avoiding the problem of overlapping boxes.

[0081] See Figure 8 As shown, Figure 8 This is a schematic diagram of a training method for a target detection model provided in another embodiment of the present invention, which may include:

[0082] Step S31: Obtain the bounding boxes and anchor boxes of the training image samples and determine the preliminary recognition category of the anchor boxes based on the bounding boxes;

[0083] Step S32: When the number of bounding boxes whose intersection-union ratio with the anchor box is greater than the preset intersection-union ratio threshold is greater than 1, the final identification category of the anchor box is determined to be the target negative sample.

[0084] Step S33: When the number of annotation boxes with an intersection-union ratio greater than the preset intersection-union ratio threshold is 1, determine whether the center of the anchor box is in the preset center area of ​​the annotation box.

[0085] Step S34: When the center of the anchor box is in the preset center area of ​​the annotation box, the final recognition category of the anchor box is determined to be the target positive sample; when the center of the anchor box is not in the preset center area of ​​the annotation box, the final recognition category of the anchor box is determined to be the target negative sample.

[0086] Step S35: Train the target detection model based at least on the anchor boxes, the final recognition category of the anchor boxes, and the labeled boxes to obtain a trained target detection model.

[0087] Steps S31 and S35 can be implemented in the same way as steps S21 and S23, and will not be described in detail here for the sake of brevity.

[0088] This method can filter anchor boxes that satisfy an intersection-union ratio (IU) greater than a preset IU threshold with multiple labeled boxes from anchor boxes initially identified as positive samples, and correct the initial identification category of the anchor box; or, when the number of labeled boxes with an IU greater than the preset IU threshold is 1 and the center of the anchor box is not in the preset center area of ​​the labeled box, the initial identification category of the anchor box is corrected. This can remove some anchor boxes that are prone to causing duplicate boxes from the anchor boxes initially identified as positive samples, which is beneficial for efficient training of the object detection model and improving the detection accuracy of the trained object detection model. In turn, it can effectively avoid the problem of duplicate boxes when performing object detection based on the trained object detection model in the future.

[0089] In some embodiments, before performing step S33, at least one of the following methods (one to three) can be used to construct the preset central region:

[0090] Method 1

[0091] Get the border length and border width of the annotation box;

[0092] Using the center point of the annotation box as the center, and the border length of the first ratio as the side length and the border width of the second ratio as the side width, construct a preset central area of ​​a rectangle, where the first ratio and the second ratio are less than 1.

[0093] Method 2

[0094] Construct a pre-defined circular central region with the center point of the annotation box as the center and the third ratio of the minimum side length of the annotation box as the radius, where the third ratio is less than 1.

[0095] Method 3

[0096] Using the center point of the annotation box as the center, and the fourth ratio of the longest side of the annotation box as the first axis and the fifth ratio of the shortest side of the annotation box as the second axis, construct a preset central region of an ellipse, where the fourth and fifth ratios are less than 1.

[0097] The above describes an object detection method provided by an embodiment of the present invention. During the training process of the object detection model, a preliminary identification category of the anchor boxes is obtained based on the bounding boxes and anchor boxes of the training image samples. For anchor boxes whose preliminary identification category is positive, the final identification category is determined based on the number of bounding boxes whose intersection-union ratio (IU) with the anchor box is greater than a preset IU threshold, or based on the number of bounding boxes whose IU with the anchor box is greater than the preset IU threshold and the positional relationship between the anchor box and the bounding boxes. At least based on the anchor boxes, the final identification category of the anchor boxes, and the bounding boxes, the object detection model to be trained is trained to obtain a trained object detection model. The present invention, based on this trained object detection model, can effectively improve the detection effect and avoid the problem of duplicate bounding boxes.

[0098] In another aspect, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the target detection method of any of the above embodiments. The computer-readable storage medium may be a storage device comprising various electronic devices; optionally, in the embodiments of the present invention, the computer-readable storage medium is a non-transitory computer-readable storage medium.

[0099] In another aspect, the present invention provides a driving device, which may include: a driving device body, a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements the target detection method in any of the above embodiments.

[0100] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A target detection method, characterized in that, include: Acquire the image to be detected; The image to be detected is input into the trained target detection model to obtain the target detection result; The object detection model is trained based on the following steps: Obtain the bounding boxes and anchor boxes of the training image samples, and determine the preliminary recognition category of the anchor boxes based on the bounding boxes; For the anchor box whose initial identification category is positive, the final identification category of the anchor box whose initial identification category is positive is determined based on the number of bounding boxes whose intersection-union ratio with the anchor box is greater than a preset intersection-union ratio threshold, or based on the number of bounding boxes whose intersection-union ratio with the anchor box is greater than a preset intersection-union ratio threshold and the positional relationship between the anchor box and the bounding boxes, wherein the final identification category includes target positive sample or target negative sample; The target detection model is trained based at least on the anchor boxes, the final recognition category of the anchor boxes, and the labeled boxes to obtain a trained target detection model.

2. The method according to claim 1, characterized in that, The determination of the final identification category of the anchor boxes whose initial identification category is positive based on the number of bounding boxes whose intersection-union ratio with the anchor box is greater than a preset intersection-union ratio threshold includes: When the number of bounding boxes whose intersection-union ratio with the anchor box is greater than a preset intersection-union ratio threshold is greater than 1, the final identification category of the anchor box is determined to be the target negative sample.

3. The method according to claim 1, characterized in that, The determination of the final identification category of the anchor boxes whose initial identification category is positive, based on the number of bounding boxes whose intersection-union ratio (IU) with the anchor boxes is greater than a preset IU threshold and the positional relationship between the anchor boxes and the bounding boxes, includes: The final identification category of the anchor box whose initial identification category is positive is determined based on the number of bounding boxes whose intersection-union ratio with the anchor box is greater than a preset intersection-union ratio threshold, and the positional relationship between the center of the anchor box and the preset center region of the bounding box.

4. The method according to claim 3, characterized in that, The determination of the final identification category of the anchor boxes whose initial identification category is positive based on the number of bounding boxes whose intersection-union ratio (IU) with the anchor boxes is greater than a preset IU threshold, and the positional relationship between the center of the anchor boxes and the preset center region of the bounding boxes, includes: When the number of annotation boxes with an intersection-union ratio greater than a preset intersection-union ratio threshold is 1, it is determined whether the center of the anchor box is in the preset center region of the annotation box. When the center of the anchor box is within the preset center area of ​​the annotation box, the final identification category of the anchor box is determined to be a target positive sample; when the center of the anchor box is not within the preset center area of ​​the annotation box, the final identification category of the anchor box is determined to be a target negative sample.

5. The method according to claim 1, characterized in that, The step of training the target detection model based at least on the anchor boxes, the final recognition category of the anchor boxes, and the labeled boxes includes: Based on the anchor boxes that are ultimately identified as positive target samples and / or negative target samples, calculate the loss function between the anchor boxes and the labeled boxes, and train the target detection model to be trained according to the loss function; or, Based on the anchor boxes whose final identification category is positive target sample and / or negative target sample, and the anchor boxes whose initial identification category is negative sample, a loss function is calculated between the anchor boxes and the labeled boxes, so as to train the target detection model to be trained according to the loss function.

6. The method according to claim 3 or 4, characterized in that, The preset central region is constructed through the following steps: Obtain the border length and border width of the annotation box; With the center point of the annotation box as the center, and the border length at a first ratio as the side length and the border width at a second ratio as the side width, a rectangular preset central region is constructed, wherein the first ratio and the second ratio are less than 1.

7. The method according to claim 3 or 4, characterized in that, The preset central region is constructed through the following steps: A pre-defined circular central region is constructed with the center point of the annotation box as the center and the third ratio of the minimum side length of the annotation box as the radius, wherein the third ratio is less than 1.

8. The method according to claim 3 or 4, characterized in that, The preset central region is constructed through the following steps: With the center point of the annotation box as the center, and the fourth ratio of the longest side length of the annotation box as the first axis and the fifth ratio of the shortest side length of the annotation box as the second axis, an elliptical preset central region is constructed, wherein the fourth ratio and the fifth ratio are less than 1.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the target detection method according to any one of claims 1 to 8.

10. A driving device, characterized in that, The device includes a driving equipment body, a memory, and a processor. The memory stores a computer program, which, when executed by the processor, implements the target detection method according to any one of claims 1 to 8.

Images