Target detection method, system, device and medium

Abstract

The invention discloses a target detection method and system, computer equipment and a medium. The method comprises the following steps: acquiring a to-be-recognized image and calculating to obtain a plurality of feature maps of the image; performing side edge connection on the plurality of feature maps to obtain a plurality of fusion feature maps, and screening out a plurality of fusion feature maps from the fusion feature maps; and respectively inputting the plurality of fused feature maps into different detection modules of a target detection model so as to respectively detect targets with different sizes. According to the scheme provided by the invention, multiple detection functions are integrated into different modules in one model, so that the number of required detection models is reduced, and the complexity and maintenance cost of model deployment are reduced.

Description

Technical Field

[0001] The present invention relates to the field of image target detection, and in particular to a target detection method, system, equipment and medium. Background Art

[0002] License plate detection and character recognition are widely used in the field of intelligent transportation. The detection and recognition of traffic participants are very important for the safe driving of autonomous vehicles. Since the proportion of license plates and vehicle pixels in the image is very different, if a single model is used to directly detect both license plates and traffic participants at the same time, it is difficult to take into account the characteristics of both. During the optimization process, the model often focuses on the detection effect of a certain type of target, resulting in the sacrifice of the accuracy of another type of target. For example, if the model is more inclined to accurately detect large-sized vehicles, then there may be more missed detections or false detections when detecting small-sized license plates, and vice versa.

[0003] In view of the above problems faced by using a single model, in actual applications, the common practice is to use two independent models to detect license plate detection and traffic participant detection respectively. The advantage of this is that each model can be optimized according to the target characteristics to be detected (such as the small size, specific shape and character distribution characteristics of the license plate, and the different types, sizes and behavior patterns of traffic participants, etc.), so that a higher accuracy rate can be achieved in the respective detection tasks, and the needs of the intelligent transportation field for license plate detection and traffic participant detection can be better met.

[0004] Although dual-model detection has many advantages, the excessive use of independent models also brings some problems in the actual deployment and maintenance process. Different models may have different dependency libraries, operating environment requirements, etc., which increases the complexity of deployment and requires more time and manpower to ensure that each model can run normally in the target system. During the use of the model, it needs to be maintained continuously. Multiple models mean that these maintenance tasks need to be carried out separately for each model, which undoubtedly increases the cost of model maintenance. Summary of the invention

[0005] In view of this, in order to overcome at least one aspect of the above problems, an embodiment of the present invention provides a target detection method, comprising the following steps: Acquire an image to be identified and calculate multiple feature maps of the image; Performing side connections on the multiple feature maps to obtain multiple fused feature maps, and selecting a number of fused feature maps from the fused feature maps; The several fused feature maps are respectively input into different detection modules of the target detection model to detect targets of different sizes respectively.

[0006] In some embodiments, performing side connections on the multiple feature maps to obtain multiple fused feature maps further includes: The Nth feature map is used for convolution calculation to obtain the Nth fused feature map; Upsample the Nth fused feature map to obtain the N-1th intermediate feature map of the same size as the N-1th feature map; Multiply the first weight by the N-1th feature map and multiply the second weight by the N-1th intermediate feature map, and add the two results to obtain the N-1th fused feature map, wherein the sum of the first weight and the second weight is 1, and the first weight is greater than the second weight.

[0007] In some embodiments, the plurality of fused feature maps are respectively input into different detection modules of the target detection model to respectively detect targets of different sizes, further comprising: Inputting the N-1th fused feature map as a first feature map into the first detection module, wherein the first detection module includes a multi-layer perceptron; The first feature map is calculated using the multi-layer perceptron to obtain the type and bounding box of each first traffic participant.

[0008] In some embodiments, the plurality of fused feature maps are respectively input into different detection modules of the target detection model to respectively detect targets of different sizes, further comprising: Inputting the second fused feature map as a second feature map into the second detection module, wherein the second detection module includes a first encoder, a first decoder, and a multi-layer perceptron; Using a first encoder to encode the second feature map to obtain a coding feature; Decoding the encoded feature using a first decoder to obtain a decoded feature; The decoded features are calculated using a multi-layer perceptron to obtain the type and bounding box of each first traffic participant.

[0009] In some embodiments, inputting the second fused feature map as the second feature map into the second detection module further includes: Calculate a position encoding vector corresponding to each point in the second feature map; Adding the vector corresponding to each point in the second feature map to the corresponding position encoding vector to obtain a second feature map with position encoding; The second feature map with position encoding is input into the second detection module.

[0010] In some embodiments, the plurality of fused feature maps are respectively input into different detection modules of the target detection model to respectively detect targets of different sizes, further comprising: Inputting the first fused feature map as a third feature map into the third detection module, wherein the third detection module includes a second encoder, a second decoder, and a plurality of feedforward neural networks; Using the second encoder to encode the third feature map to obtain a coding feature; Inputting the encoded features and a plurality of pre-learned vectors into the second decoder for decoding to obtain a plurality of output results; Each of the feedforward neural networks is used to calculate one of the output results to obtain the detection result of each other target.

[0011] In some embodiments, the first fused feature map is input as a third feature map to the third detection module, further comprising: Calculate the position encoding vector corresponding to each point in the third feature map; Adding the vector corresponding to each point in the third feature map to the corresponding position encoding vector to obtain a third feature map with position encoding; The third feature map with position encoding is input into the third detection module.

[0012] Based on the same inventive concept, according to another aspect of the present invention, an embodiment of the present invention further provides a target detection system, including: A calculation module, configured to obtain an image to be identified and calculate a plurality of feature maps of the image; A fusion module is configured to perform side connections on the multiple feature maps to obtain multiple fused feature maps, and to select a plurality of fused feature maps from the fused feature maps; The detection module is configured to input the plurality of fused feature maps into different detection modules of the target detection model respectively, so as to detect targets of different sizes respectively.

[0013] Based on the same inventive concept, according to another aspect of the present invention, an embodiment of the present invention further provides a computer device, including: at least one processor; and A memory, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the program, the steps of any one of the target detection methods described above are performed.

[0014] Based on the same inventive concept, according to another aspect of the present invention, an embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of any one of the target detection methods described above are performed.

[0015] The present invention has one of the following beneficial technical effects: the solution proposed by the present invention deploys multiple modules in one model, and each module is set to detect traffic participants of different sizes. For example, the first detection module detects first traffic participants such as ambulances, buses, and fire trucks, the second detection module detects second traffic participants such as bicycles and pedestrians, and the third detection module detects other targets such as license plates. Such division of labor enables the model to carry out targeted processing of targets of various sizes, ensuring that the detection of targets of different sizes has a high accuracy rate. Moreover, by integrating multiple detection functions into different modules in one model, the present application eliminates the need to develop and deploy a complete model for each target of different sizes, thereby reducing the number of required detection models. At the same time, only one model needs to be deployed for related operations, further reducing the complexity and maintenance cost of model deployment. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other embodiments can be obtained based on these drawings without paying creative work.

[0017] Figure 1 A schematic diagram of a flow chart of a target detection method provided by an embodiment of the present invention; Figure 2 A flowchart of a target detection method provided by an embodiment of the present invention; Figure 3 A schematic diagram of the structure of a target detection system provided by an embodiment of the present invention; Figure 4 A schematic diagram of the structure of a computer device provided by an embodiment of the present invention; Figure 5 A schematic diagram of the structure of a computer-readable storage medium provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the embodiments of the present invention are further described in detail below in combination with specific embodiments and with reference to the accompanying drawings.

[0019] It should be noted that all expressions using "first" and "second" in the embodiments of the present invention are for distinguishing two non-identical entities with the same name or non-identical parameters. It can be seen that "first" and "second" are only for the convenience of expression and should not be understood as limitations on the embodiments of the present invention. The subsequent embodiments will not explain this one by one.

[0020] According to one aspect of the present invention, an embodiment of the present invention provides a target detection method, such as Figure 1 As shown, it may include the steps of: S1, obtaining an image to be recognized and calculating a plurality of feature maps of the image; S2, performing side connections on the multiple feature maps to obtain multiple fused feature maps, and selecting a number of fused feature maps from the fused feature maps; S3, inputting the plurality of fused feature maps into different detection modules of the target detection model respectively, so as to detect targets of different sizes respectively.

[0021] The solution proposed by the present invention deploys multiple modules in one model, and each module is set to detect traffic participants of different sizes. For example, the first detection module detects first traffic participants such as ambulances, buses, and fire trucks, the second detection module detects second traffic participants such as bicycles and pedestrians, and the third detection module detects other targets such as license plates. Such division of labor enables the model to carry out targeted processing of targets of various sizes, ensuring that the detection of targets of different sizes has a high accuracy rate. In addition, by integrating multiple detection functions into different modules in one model, the present application eliminates the need to develop and deploy a complete model for each target of different sizes, thereby reducing the number of required detection models. At the same time, only one model needs to be deployed for related operations, further reducing the complexity and maintenance cost of model deployment.

[0022] In some embodiments, step S1, obtains an image to be identified and calculates multiple feature maps of the image. Specifically, after the image is collected, for example, after the image is collected by a roadside image sensor, a convolutional neural network (CNN) can be used to perform convolution calculations on the image to obtain multiple feature maps, wherein the multiple feature maps include shallow feature maps and deep feature maps. The shallow feature map refers to the features extracted by the convolution layer close to the input layer, which mainly captures the local texture and simple geometric structure of the image; the deep feature map refers to the features generated by the convolution layer in the deeper layer of the network, which has higher abstractness.

[0023] For example, Figure 2As shown in FIG. 1 , in module A, a convolutional neural network is used to perform convolution calculation on the input image to obtain four feature maps, namely the first feature map d1, the second feature map c1, the third feature map b1 and the fourth feature map a1. Since the first feature map d1 and the second feature map c1 are closer to the convolution layer of the input layer, they can be identified as shallow feature maps; the third feature map b1 and the fourth feature map a1 are features generated by the deeper convolution layer of the network and can be identified as deep feature maps.

[0024] It should be noted that the number of convolutional layers of a convolutional neural network can be determined according to actual needs, and parameters can be optimized through training and learning. For example, a large amount of labeled image data (training set) is input into the network for forward propagation, and the error (loss function) between the predicted output and the true label is calculated. This error will be propagated from the output layer to the input layer through the back propagation algorithm, and the parameters in each convolutional layer will be adjusted according to the error.

[0025] In some embodiments, S2, performing side connections on the multiple feature maps to obtain multiple fused feature maps, further comprises: The Nth feature map is used for convolution calculation to obtain the Nth fused feature map; Upsample the Nth fused feature map to obtain the N-1th intermediate feature map of the same size as the N-1th feature map; Multiply the first weight by the N-1th feature map and multiply the second weight by the N-1th intermediate feature map, and add the two results to obtain the N-1th fused feature map, wherein the sum of the first weight and the second weight is 1, and the first weight is greater than the second weight.

[0026] Specifically, after using Convolutional Neural Networks (CNN) to perform convolution calculations on an image to obtain multiple feature maps, the shallow feature map and the deep feature map can be connected from top to bottom, and feature maps of different scales can be fused so that each layer of feature map has deep and shallow semantic information. At the same time, different weights are assigned to the deep feature map and the shallow feature map during fusion, and the weight of the deep feature map is smaller than that of the shallow feature map. In this way, by assigning different weights instead of using a unified deep and shallow fusion ratio, the background interference can be reduced during small target detection, and the accuracy of small target detection can be improved. For example, since the license plate area occupies a small proportion of pixels, too much deep semantic information contains more background feature information, which will reduce the accuracy of license plate detection. Therefore, reducing the weight of the deep feature map during fusion can improve the accuracy of small target detection. In some embodiments, the sum of the first weight and the second weight is equal to 1, and the first weight is greater than 0.8.

[0027] For example, Figure 2 As shown, in module A, after obtaining the first feature map d1, the second feature map c1, the third feature map b1 and the fourth feature map a1, the fourth feature map a1 is convolved using a 1×1 convolution layer to obtain the fourth fused feature map. Then, based on the fourth fused feature map a2 and the third feature map b1, the third fused feature map b2 is obtained, that is, the fourth fused feature map a2 is downsampled to obtain the third intermediate feature map a2 of the same size as the third feature map b1. 下 , using the first weight P1 and the second weight P2 to calculate: b1×P1+ a2 下 ×P2, we get the third fused feature map b2. Similarly, using c1×P1+ b2 下 ×P2 to get the second fused feature map c2, using d1×P1+ c2 下 ×P2 gets the first fused feature map d2. In this way, module A outputs four fused feature maps, which are the fourth fused feature map a2, the third fused feature map b2, the second fused feature map c2, and the first fused feature map d2 from top to bottom.

[0028] In some embodiments, the plurality of fused feature maps are respectively input into different detection modules of the target detection model to respectively detect targets of different sizes, further comprising: Inputting the N-1th fused feature map as a first feature map into the first detection module, wherein the first detection module includes a multi-layer perceptron; The first feature map is calculated using the multi-layer perceptron to obtain the type and bounding box of each first traffic participant.

[0029] Specifically, the first detection module mainly detects first traffic participants such as ambulances, buses, and fire trucks, and may include a multilayer perceptron (MLP). In the first detection module, the type and bounding box (bbox) of each first traffic participant can be obtained by calculating the input first feature map using a multilayer perceptron. The bounding box is usually a rectangular box that is used to define the specific position and range of the first traffic participant in the image. In some embodiments, the multilayer perceptron can optimize parameters through training and learning, and since the first detection module is only responsible for detecting first traffic participants of larger sizes, more precise parameter adjustments can be made during training and learning according to the characteristics of the target size, so that the final detection result is more accurate.

[0030] For example, Figure 2As shown, since four fused feature maps are output in module A, the third fused feature map b2 is input into the first detection module B as the first feature map, and a multilayer perceptron (MLP) is used in module B to calculate and obtain the detection result of the motor vehicle and the corresponding vehicle category.

[0031] In some embodiments, the plurality of fused feature maps are respectively input into different detection modules of the target detection model to respectively detect targets of different sizes, further comprising: Inputting the second fused feature map as a second feature map into the second detection module, wherein the second detection module includes a first encoder, a first decoder, and a multi-layer perceptron; Using a first encoder to encode the second feature map to obtain a coding feature; Decoding the encoded feature using a first decoder to obtain a decoded feature; The decoded features are calculated using a multi-layer perceptron to obtain the type and bounding box of each first traffic participant.

[0032] Specifically, the second detection module mainly detects some small-sized second traffic participants such as pedestrians, motorcycles, and electric vehicles, and may include a first encoder, a first decoder, and a multi-layer perceptron. In the second detection module, the first encoder calculates the input second feature map to obtain global context information, and then inputs the global context information into the first decoder for decoding. The decoding result is finally input into the multi-layer perceptron to calculate the detection result bbox and the corresponding category of the second traffic participant. In some embodiments, the first encoder, the first decoder, and the multi-layer perceptron can optimize parameters through training and learning, and since the second detection module is only responsible for detecting small-sized second traffic participants, more precise parameter adjustments can be made during training and learning according to the characteristics of the target size, so that the final detection result is more accurate.

[0033] For example, Figure 2 As shown, since four fused feature maps are output in module A, the second fused feature map c2 is input as the second feature map into the second detection module C. In module C, the first encoder, the first decoder and the multi-layer perceptron are used to calculate and obtain the type and bounding box of the second traffic participants such as bicycles and pedestrians.

[0034] In some embodiments, inputting the second fused feature map as the second feature map into the second detection module further includes: Calculate a position encoding vector corresponding to each point in the second feature map; Adding the vector corresponding to each point in the second feature map to the corresponding position encoding vector to obtain a second feature map with position encoding; The second feature map with position encoding is input into the second detection module.

[0035] Specifically, sine function and cosine function can be used for position encoding, that is, using PE （pos，2i） =sin( pos / 10000 2i / d ), PE （pos，2i+1） =cos( pos / 10000 2i / d ) calculates the position code, where pos takes values ​​of 1 to n, n is the number of feature points in the second feature map, i represents the dimension, if d is an even number, i takes values ​​of 0 to d / 2, if i is an odd number, i takes values ​​of 0 to (d+1) / 2, d is the preset dimension of the feature point. In this way, by adding position codes to the feature map, the specific position of the feature in the image can be better known, so that the position information can be fully utilized in subsequent calculations to improve the detection accuracy.

[0036] For example, if the dimension of the second feature map before encoding is 512 and the number of position points is 1024, then pos takes values ​​of 1 to 1024 and i takes values ​​of 0 to 255. When the above formula is used to calculate the position encoding, when pos takes 1, i takes values ​​of 0 to 255, and the position vector of the 2i position uses PE （pos，2i） Indicates that the position vector at position 2i+1 uses PE （pos，2i+1） Indicates that the position vector is expressed as [sin ( 1 / 10000 0 / 512 ), cos( 1 / 10000 1 / 512 ), sin( 1 / 10000 2 / 512 ), cos( 1 / 10000 3 / 512 )...sin( 1 / 10000 511 / 512 ), cos( 1 / 10000 512 / 512 )], that is, the 512-dimensional position encoding vector corresponding to the first feature point. When pos is 2, the position vector of token2 is expressed as [sin ( 2 / 10000 0 / 512 ), cos( 2 / 10000 1 / 512 ), sin( 2 / 10000 2 / 512 ), cos( 2 / 100003 / 512 )...sin( 2 / 10000 511 / 512 ), cos( 2 / 10000 512 / 512 )], and so on. After calculation by the above formula, each feature point will get a 512-dimensional position encoding vector, thus obtaining a 1024×512 position encoding vector. Then the 1024×512 position encoding vector is added to the corresponding feature map, that is, the vector of each feature point is added to the corresponding position encoding vector, thereby obtaining a second feature map that integrates the position information.

[0037] In some embodiments, the plurality of fused feature maps are respectively input into different detection modules of the target detection model to respectively detect targets of different sizes, further comprising: Inputting the first fused feature map as a third feature map into the third detection module, wherein the third detection module includes a second encoder, a second decoder, and a plurality of feedforward neural networks; Using the second encoder to encode the third feature map to obtain a coding feature; Inputting the encoded features and a plurality of pre-learned vectors into the second decoder for decoding to obtain a plurality of output results; Each of the feedforward neural networks is used to calculate one of the output results to obtain the detection result of each other target.

[0038] Specifically, the third detection module mainly detects small targets such as license plates, and may include a second encoder, a second decoder, and multiple feedforward neural networks (FFN). In the third detection module, the second encoder calculates the input third feature map to obtain global context information, and then the global context information and multiple pre-learned vectors (Object queries) are input into the second decoder for decoding calculation, that is, each vector is calculated with the global context information to obtain multiple calculation results, and each calculation result is input into a feedforward neural network, and the feedforward neural network is calculated to obtain a bounding box and a corresponding category. In some embodiments, the second encoder, the second decoder, multiple vectors, and multiple feedforward neural networks can be optimized by training and learning, wherein the number of trained and learned vectors and feedforward neural networks can be determined according to the number of detection targets. For example, when a small target such as a license plate needs to be detected, one vector and one feedforward neural network can be trained and learned, and when three small targets need to be detected, three vectors and three feedforward neural networks can be trained and learned. Moreover, since the third detection module is only responsible for detecting other smaller targets, compared with the second detection module directly using the multi-layer perceptron for detection, the third detection module combines vectors to detect small targets, with higher precision and more accurate detection results.

[0039] For example, Figure 2 As shown, since four fused feature maps are output in module A, the first fused feature map d2 is input into the third detection module D as the third feature map. In module D, the second encoder, the second decoder, two vectors and two feedforward neural networks (FFN) are used to calculate the license plate type (single-layer license plate and double-layer license plate) and bbox. Finally, the license plate detection result is input into OCR (Optical Character Recognition) for license plate character recognition to obtain the license plate recognition result.

[0040] In some embodiments, the first fused feature map is input as a third feature map to the third detection module, further comprising: Calculate the position encoding vector corresponding to each point in the third feature map; Adding the vector corresponding to each point in the third feature map to the corresponding position encoding vector to obtain a third feature map with position encoding; The third feature map with position encoding is input into the third detection module.

[0041] Specifically, sine function and cosine function can be used for position encoding, that is, using PE （pos，2i） =sin( pos / 10000 2i / d ), PE （pos，2i+1） =cos( pos / 10000 2i / d ) calculates the position code, where pos takes values ​​of 1 to n, n is the number of feature points in the third feature map, i represents the dimension, if d is an even number, i takes values ​​of 0 to d / 2, if i is an odd number, i takes values ​​of 0 to (d+1) / 2, and d is the preset dimension of the feature point. In this way, by adding position codes to the feature map, the specific position of the feature in the image can be better known, so that the position information can be fully utilized in subsequent calculations to improve the detection accuracy.

[0042] For example, if the dimension of the third feature map before encoding is 512 and the number of position points is 1024, then pos takes values ​​of 1 to 1024 and i takes values ​​of 0 to 255. When the above formula is used to calculate the position encoding, when pos takes 1, i takes values ​​of 0 to 255, and the position vector of the 2i position uses PE （pos，2i） Indicates that the position vector at position 2i+1 uses PE （pos，2i+1） Indicates that the position vector is expressed as [sin ( 1 / 10000 0 / 512 ), cos( 1 / 10000 1 / 512 ), sin( 1 / 10000 2 / 512 ), cos( 1 / 10000 3 / 512 )...sin( 1 / 10000 511 / 512 ), cos( 1 / 10000 512 / 512 )], that is, the 512-dimensional position encoding vector corresponding to the first feature point. When pos is 2, the position vector of token2 is expressed as [sin ( 2 / 10000 0 / 512 ), cos( 2 / 10000 1 / 512 ), sin( 2 / 10000 2 / 512 ), cos( 2 / 10000 3 / 512 )...sin( 2 / 10000 511 / 512 ), cos( 2 / 10000 512 / 512)], and so on. After calculation by the above formula, each feature point will get a 512-dimensional position encoding vector, thus obtaining a 1024×512 position encoding vector. Then the 1024×512 position encoding vector is added to the corresponding feature map, that is, the vector of each feature point is added to the corresponding position encoding vector, thereby obtaining a third feature map that integrates the position information.

[0043] The solution proposed by the present invention deploys multiple modules in one model, and each module is set to detect traffic participants of different sizes. For example, the first detection module detects first traffic participants such as ambulances, buses, and fire trucks, the second detection module detects second traffic participants such as bicycles and pedestrians, and the third detection module detects other targets such as license plates. Such division of labor enables the model to carry out targeted processing of targets of various sizes, ensuring that the detection of targets of different sizes has a high accuracy rate. In addition, by integrating multiple detection functions into different modules in one model, the present application eliminates the need to develop and deploy a complete model for each target of different sizes, thereby reducing the number of required detection models. At the same time, only one model needs to be deployed for related operations, further reducing the complexity and maintenance cost of model deployment.

[0044] Based on the same inventive concept, according to another aspect of the present invention, an embodiment of the present invention further provides a target detection system 400, such as Figure 3 As shown, including: A calculation module 401 is configured to obtain an image to be identified and calculate a plurality of feature maps of the image; A fusion module 402 is configured to perform side connections on the multiple feature maps to obtain multiple fused feature maps, and select a number of fused feature maps from the fused feature maps; The detection module 403 is configured to input the plurality of fused feature maps into different detection modules of the target detection model respectively, so as to detect targets of different sizes respectively.

[0045] Based on the same inventive concept, according to another aspect of the present invention, Figure 4 As shown, an embodiment of the present invention further provides a computer device 501, including: at least one processor 520; and The memory 510 stores a computer program 511 that can be run on the processor. When the processor 520 executes the program, the steps of any one of the above target detection methods are performed.

[0046] Based on the same inventive concept, according to another aspect of the present invention, Figure 5As shown, an embodiment of the present invention further provides a computer-readable storage medium 601, which stores a computer program 610. When the computer program 610 is executed by a processor, the steps of any of the above target detection methods are performed.

[0047] Finally, it should be noted that a person skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods.

[0048] Furthermore, it should be appreciated that the computer-readable storage medium (eg, memory) herein may be either volatile memory or nonvolatile memory, or may include both volatile memory and nonvolatile memory.

[0049] It will also be appreciated by those skilled in the art that various exemplary logic blocks, modules, circuits and algorithm steps described in conjunction with the disclosure herein can be implemented as electronic hardware, computer software or a combination of the two. In order to clearly illustrate this interchangeability of hardware and software, a general description has been given to the functions of various schematic components, blocks, modules, circuits and steps. Whether this function is implemented as software or hardware depends on specific applications and the design constraints imposed on the entire system. Those skilled in the art can implement the function in various ways for each specific application, but this implementation decision should not be interpreted as causing a departure from the disclosed scope of the embodiments of the present invention.

[0050] The above are exemplary embodiments disclosed in the present invention, but it should be noted that various changes and modifications may be made without departing from the scope disclosed in the embodiments of the present invention as defined in the claims. The functions, steps and / or actions of the method claims according to the disclosed embodiments described herein do not need to be performed in any particular order. In addition, although the elements disclosed in the embodiments of the present invention may be described or required in individual form, they may also be understood as multiple unless explicitly limited to the singular.

[0051] It should be understood that, as used herein, the singular forms "a", "an" are intended to include the plural forms as well, unless the context clearly supports an exception. It should also be understood that, as used herein, "and / or" refers to any and all possible combinations including one or more of the associated listed items.

[0052] The serial numbers of the embodiments disclosed in the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0053] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0054] A person skilled in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the disclosure of the embodiments of the present invention (including the claims) is limited to these examples; under the concept of the embodiments of the present invention, the technical features in the above embodiments or different embodiments can also be combined, and there are many other changes in different aspects of the above embodiments of the present invention, which are not provided in detail for the sake of simplicity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present invention should be included in the protection scope of the embodiments of the present invention.

Claims

1. A target detection method, characterized in that: The following steps are involved: Acquire an image to be identified and calculate multiple feature maps of the image; Performing side connections on the multiple feature maps to obtain multiple fused feature maps, and selecting a number of fused feature maps from the fused feature maps; The several fused feature maps are respectively input into different detection modules of the target detection model to detect targets of different sizes respectively.

2. The method according to claim 1, characterized in that Performing side connections on the multiple feature maps to obtain multiple fused feature maps further includes: The Nth feature map is used for convolution calculation to obtain the Nth fused feature map; Upsample the Nth fused feature map to obtain the N-1th intermediate feature map of the same size as the N-1th feature map; Multiply the first weight by the N-1th feature map and multiply the second weight by the N-1th intermediate feature map, and add the two results to obtain the N-1th fused feature map, wherein the sum of the first weight and the second weight is 1, and the first weight is greater than the second weight.

3. The method according to claim 2, characterized in that Inputting the plurality of fused feature maps into different detection modules of the target detection model respectively to detect targets of different sizes respectively, further comprising: Inputting the N-1th fused feature map as a first feature map into a first detection module, wherein the first detection module includes a multi-layer perceptron; The first feature map is calculated using the multi-layer perceptron to obtain the type and bounding box of each first traffic participant.

4. The method according to claim 2, characterized in that Inputting the plurality of fused feature maps into different detection modules of the target detection model respectively to detect targets of different sizes respectively, further comprising: Inputting the second fused feature map as a second feature map into a second detection module, wherein the second detection module includes a first encoder, a first decoder, and a multi-layer perceptron; Using a first encoder to encode the second feature map to obtain a coding feature; Decoding the encoded feature using a first decoder to obtain a decoded feature; The decoded features are calculated using a multi-layer perceptron to obtain the type and bounding box of each first traffic participant.

5. The method according to claim 4, characterized in that Inputting the second fused feature map as the second feature map into the second detection module further includes: Calculate a position encoding vector corresponding to each point in the second feature map; Adding the vector corresponding to each point in the second feature map to the corresponding position encoding vector to obtain a second feature map with position encoding; The second feature map with position encoding is input into the second detection module.

6. The method according to claim 2, characterized in that Inputting the plurality of fused feature maps into different detection modules of the target detection model respectively to detect targets of different sizes respectively, further comprising: Inputting the first fused feature map as a third feature map into a third detection module, wherein the third detection module includes a second encoder, a second decoder, and a plurality of feedforward neural networks; Using the second encoder to encode the third feature map to obtain a coding feature; Inputting the encoded features and a plurality of pre-learned vectors into the second decoder for decoding to obtain a plurality of output results; Each of the feedforward neural networks is used to calculate one of the output results to obtain the detection result of each other target.

7. The method according to claim 6, characterized in that Inputting the first fused feature map as a third feature map into a third detection module, further comprising: Calculate the position encoding vector corresponding to each point in the third feature map; Adding the vector corresponding to each point in the third feature map to the corresponding position encoding vector to obtain a third feature map with position encoding; The third feature map with position encoding is input into the third detection module.

8. A target detection system, characterized in that: include: A calculation module, configured to obtain an image to be identified and calculate a plurality of feature maps of the image; A fusion module is configured to perform side connections on the multiple feature maps to obtain multiple fused feature maps, and to select a plurality of fused feature maps from the fused feature maps; The detection module is configured to input the plurality of fused feature maps into different detection modules of the target detection model respectively, so as to detect targets of different sizes respectively.

9. A computer device comprising: at least one processor; as well as A memory storing a computer program executable on the processor, wherein the processor executes the steps of the method according to any one of claims 1 to 7 when executing the program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are performed.

Images