An occluded pedestrian detection method based on feature focusing and information fusion

Abstract

The application provides a kind of based on feature focusing and information fusion's occluded pedestrian detection method, the method first with the high-resolution network HRNet of convolutional neural network as main network, feature extraction is carried out to input image, so that high-resolution features are always retained while extracting image feature information;Secondly, through the dynamic adjustment of the importance of each channel and spatial position of feature map by feature focusing module, the key region feature information is strengthened and the background noise is suppressed, so that the main network focuses on the target area;Then, the hierarchical feature fusion mechanism efficiently fuses high-level semantic information and location detail information;Finally, the fused feature map is passed through the detection head to generate the center point feature map, the scale feature map and the center point offset feature map, and the feature map is interpreted as a detection result, and the pedestrian region with a bounding box is output;The application can solve the occlusion problem in pedestrian detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pedestrian detection, and in particular to a method for occluded pedestrian detection based on feature focusing and information fusion. BACKGROUND

[0002] Pedestrian detection is one of the important branches of target detection tasks. In recent years, pedestrian detection has important practical application value in the fields of automatic driving, robot navigation system and intelligent monitoring, and its core purpose is to accurately identify all pedestrian targets in the image. However, pedestrians are often occluded by other pedestrians or objects, making it difficult to extract complete feature information, thereby causing the detector to miss detection. Therefore, how to effectively solve the problem of pedestrian occlusion has become a key challenge to further improve the performance of the pedestrian detector.

[0003] In recent years, researchers have mainly explored solutions from three aspects: post-processing optimization, model improvement based on specific models, and body part feature processing. First, researchers have explored post-processing optimization methods, designed specific NMS improvement strategies to punish inaccurate detection boxes, although researchers continue to optimize NMS, but due to the characteristics of suppressing sample boxes, severely occluded positive samples may still be misjudged as false positives, thereby limiting the ability to solve the occlusion problem. Second, with the great success of Transformer in the field of computer vision, researchers have begun to explore methods based on the DETR model, which successfully avoids the shortcomings of NMS by using a one-to-one label matching strategy for prediction. However, the self-attention mechanism introduced by Transformer is limited by the quadratic time complexity, making it difficult to deploy on high-resolution feature maps. Third, in view of the characteristics of the visible body parts of pedestrians under occlusion, researchers have further proposed methods based on body part models. This method uses a divide-and-conquer strategy to design corresponding detectors for different body parts to assist in overall pedestrian detection.

[0004] However, the feature information of occluded pedestrian targets is often largely missing, and the above three methods do not fully consider the difficulties faced by the network in extracting features, such as limited network attention resources, the target area usually occupies a small proportion in the overall image, and background pixels often occupy a large amount of space, making it difficult for the model to obtain enough feature information and easily disturbed by background noise. SUMMARY

[0005] The present application proposes a method for occluded pedestrian detection based on feature focusing and information fusion, which can solve the problem of occlusion in pedestrian detection.

[0006] The present application adopts the following technical solutions.

[0007] A kind of occluded pedestrian detection method based on feature focusing and information fusion, the method first with the high-resolution network HRNet of convolutional neural network as main network, the feature of input image is extracted, so as to always keep high-resolution features while extracting image feature information;Secondly, through the feature focusing module, the importance of each channel and spatial position of feature map is dynamically adjusted to strengthen key area feature information and suppress background noise, so that the main network focuses on target area;Then with hierarchical feature fusion mechanism, high-level semantic information and location detail information are efficiently fused;Finally, the feature map after fusion is passed through detection head, and center point feature map, scale feature map and center point offset feature map are generated, and the feature map is interpreted into detection result, and the pedestrian region with boundary box is output.

[0008] The method comprises the following steps:

[0009] Step S1, input image, carry out feature extraction;

[0010] Step S2, design feature focusing module, focus and enhance the feature extracted by main network;

[0011] Step S3, design hierarchical feature fusion mechanism, and the multiple different resolution feature maps output by step S2 are efficiently fused;

[0012] Step S4, the feature map is interpreted into detection result by detection head.

[0013] In step S1, the convolutional neural network HRNet is used as the main network to process the input image, and four feature maps f1, f2, f3 and f4 of different resolutions are output, and the resolutions are 1 / 4, 1 / 8, 1 / 16 and 1 / 32 times of the input image respectively.

[0014] The main network HRNet adopts parallel architecture and is composed of four branches with different resolutions;When extracting image features, the high-resolution subnetwork is used for feature extraction in the initial stage, and then the parallel subnetworks from high resolution to low resolution are gradually added, specifically in the four stages, the first stage contains only one high-resolution branch, from the second stage to the fourth stage, a parallel branch is added in each stage, the input of the next stage is derived from the output of each branch of the previous stage, and the resolution of the new branch is half of the lowest resolution branch of the previous stage, and the number of channels is doubled;The feature maps of parallel branches have scales due to different resolutions, so they can focus on targets of different scales of the original image, and the main network fuses feature maps of different resolutions through upsampling and downsampling operations, so that information exchange between different levels is realized, and the multi-scale expression ability of feature maps is improved.

[0015] When in the pedestrian detection task, if the occlusion causes the detector to misjudge or miss, relying only on the backbone network HRNet to extract pedestrian features has been difficult to meet the accuracy requirements. When the feature focusing module and the hierarchical fusion strategy are used, the features of the occluded pedestrians are enhanced through focusing and information fusion of the features.

[0016] In step S2, the feature focusing module performs fine processing on the input feature map from the dimensions of channel attention and spatial attention;

[0017] In terms of channel attention, the feature focusing module first performs global average AvgPool and maximum pooling MaxPool operations in the spatial dimension to condense two global description vectors to capture the global importance information of each channel in the entire feature map. Then, a small fully connected neural network FC is used to learn and calculate the channel importance weight for each global description vector. The running results are added and normalized by the Sigmoid function to output a weight vector representing the relative importance of each channel Finally, the weight vector acts on the input feature map f i through multiplication to obtain a refined feature map F i 1 , which realizes the reweighting of feature responses. This process improves the sensitivity of the model to key feature channels and helps the network dynamically focus on information-rich channels, enhancing the model's representation ability and detection performance.

[0018] In terms of spatial attention, the feature focusing module performs global average AvgPool and maximum pooling MaxPool in the channel dimension to focus on the potential information of the target region, then performs channel dimension concatenation Concat, further compresses the spatial information through 7x7 convolution Conv, and uses the Sigmoid function for normalization to generate a single-channel spatial attention map Finally, the spatial attention map is multiplied by the first refined feature map F i 1 to reweight the features again to obtain the second refined feature map F i 2 Through spatial fine processing, the sensitivity of the backbone network to key spatial positions is improved, focusing on the target pedestrian region and enhancing the target perception ability of the network.

[0019] Step S2 uses the feature focusing module to dynamically adjust the importance of each channel and spatial position in the feature map, and double-refines the feature information, thereby guiding the neural network model to focus on the feature information of the key region and effectively suppressing the interference of noise. The specific formula is:

[0020]

[0021] where f i are input feature maps of different scales, where (i=1, 2, 3, 4), F i 1 is f i after channel attention is the refined feature map after weighting, F i 2 is F i 1 after spatial attention is the output feature map after weighting again.

[0022] In step S3, the four different resolution feature maps output in step S2 are fused efficiently by a hierarchical feature fusion mechanism; specifically, first, the three low-resolution feature maps containing rich context information are spliced in the channel dimension for the first time to enhance the representation of deep semantic information; then, the first spliced feature map is spliced with the high-resolution feature map containing more detailed position information for the second time to realize comprehensive fusion of high-level semantic information and position detail information; the specific formula is:

[0023]

[0024] T=Concat[F1 2 ,U 1 / 4 (t)] (6)

[0025] In the above formula, and The pixel ratio of the corresponding feature map to the original input image is 1 / 4, 1 / 8, 1 / 16 and 1 / 32 times respectively;

[0026] and indicate that the feature map and is upsampled to 1 / 8 times by the bilinear interpolation algorithm, and Concat indicates splicing in the channel dimension to form the preliminary fusion feature map Then U 1 / 4 (t) is upsampled to 1 / 4 times by the bilinear interpolation algorithm, and finally spliced with the high-resolution feature map F1 2 in the channel dimension to obtain the comprehensive fusion output feature map

[0027] The specific method of step S4 is: adding a detection head to the spliced feature map for detection; the feature map for detection first passes through a 3*3 convolution kernel, and then three 1*1 convolution kernels are attached in parallel to generate a pedestrian center point feature map, a pedestrian scale feature map and a center point offset feature map; then, the pedestrian candidate box of the input image is automatically generated according to the corresponding scale in the center point feature map and the scale feature map; finally, the position of the pedestrian center point is fine-tuned through the center point offset prediction;

[0028] Therefore, the pedestrian detection can be expressed as:

[0029] Dets=H(φ det )={cls(φ det ),reg(φ det )} (7)

[0030] Wherein, H(.) represents the detection head, and φ det represents the feature map for detection; cls(.) is the predicted classification confidence, and reg(.) is the predicted scale and center point offset of the pedestrian.

[0031] In step S4, focalloss is used to train the target function L c of the pedestrian center point position prediction in the detection head:

[0032]

[0033] Wherein p ij and α ij are defined as formulas (9) and (10) respectively:

[0034]

[0035] In the formula, y ij represents the center point falling in position (i,j), p ij ∈[0,1] represents the probability that the predicted coordinate point belongs to the center, and β and γ are two hyperparameters, which are set as β=4 and γ=4. M ij is a Gaussian heat map, which aims to reduce the uncertainty of a large number of negative samples around the positive sample;

[0036] The scale prediction is a regression task using SmoothL1 loss, as shown in formula (11):

[0037]

[0038] In the formula, s k and represent the predicted value and the true value of the kth target scale;

[0039] The offset prediction is also a regression task using the SmoothL1 loss, as shown in equation (12):

[0040]

[0041] In the formula, o k And represents the predicted value and the true value of the kth target offset;

[0042] The total loss is composed of the center loss, the scale loss and the offset loss; represented as equation (13):

[0043] L = lambda c L c + lambda s L s + lambda o L o (13)

[0044] In the formula, lambda c , lambda s and lambda o are respectively set to 0.01, 1 and 0.1.

[0045] In summary, the present application firstly extracts the features of the input image through the high-resolution network (HRNet), and always retains the high-resolution features while extracting the image feature information; secondly, designs a feature focusing module, dynamically adjusts the importance of each channel and spatial position of the feature map, strengthens the key region feature information and suppresses the background noise, so that the network focuses on the target region; then, a hierarchical feature fusion mechanism is designed to efficiently fuse high-level semantic information and position detail information; finally, the fused feature map is passed through the detection head to generate the center point feature map, the scale feature map and the center point offset feature map, and the feature map is interpreted into the detection result, and the pedestrian with the bounding box is output. The feature focusing module performs fine processing on the input feature map from two dimensions of channel attention and spatial attention, realizes the double refinement of the feature information, thereby guiding the model to focus on the key region feature information and effectively suppressing the interference of noise. The hierarchical feature fusion mechanism adjusts the resolution of the low-resolution feature map using the bilinear interpolation algorithm, performs hierarchical splicing operation in the channel dimension, realizes hierarchical fusion of multi-scale feature maps, and obtains the comprehensive fusion output feature map. Finally, the refined multi-scale feature map is convolved to simplify the pedestrian detection task into a simple center point and scale prediction task; the present application can solve the occlusion problem in pedestrian detection. BRIEF DESCRIPTION OF DRAWINGS

[0046] The present application will be further described in detail below in combination with the drawings and specific embodiments:

[0047] Attached Figure 1 is a schematic diagram of the method flow of the present application;

[0048] Appendix Figure 2 This is a schematic diagram of the overall network framework of the present invention;

[0049] Appendix Figure 3 This is a schematic diagram of the feature focusing module structure of the present invention. Detailed Implementation

[0050] like Figure 1 , Figure 2 As shown, an occluded pedestrian detection method based on feature focusing and information fusion is proposed. The method first uses a high-resolution convolutional neural network (HRNet) as the backbone to extract features from the input image, ensuring high-resolution features are preserved while extracting image feature information. Secondly, a feature focusing module dynamically adjusts the importance of each channel and spatial position of the feature map to enhance key region features and suppress background noise, enabling the backbone network to focus on the target region. Then, a hierarchical feature fusion mechanism efficiently fuses high-level semantic information and location detail information. Finally, the fused feature map is passed through a detection head to generate a center point feature map, a scale feature map, and a center point offset feature map. The feature map is then interpreted into a detection result, outputting a pedestrian region with bounding boxes.

[0051] The method includes the following steps;

[0052] Step S1: Input image and perform feature extraction;

[0053] Step S2: Design a feature focusing module to focus and enhance the features extracted by the backbone network;

[0054] Step S3: Design a hierarchical feature fusion mechanism to efficiently fuse multiple feature maps of different resolutions output in step S2;

[0055] Step S4: Use the detection head to interpret the feature map into detection results.

[0056] In step S1, the convolutional neural network HRNet is used as the backbone network to process the input image and output four feature maps f1, f2, f3 and f4 with different resolutions. The ratios of their resolutions to the input image are 1 / 4, 1 / 8, 1 / 16 and 1 / 32, respectively.

[0057] The backbone network HRNet adopts a parallel architecture and is composed of four branches with different resolutions; when extracting image features, the initial stage carries out feature extraction with a high-resolution subnetwork, and subsequent stages gradually add parallel subnetworks from high resolution to low resolution; specifically, in the four stages, the first stage contains only one high-resolution branch, and from the second stage to the fourth stage, one parallel branch is added to each stage, the input of the latter stage is derived from the output of each branch of the former stage, and the resolution of the new branch is half of the lowest resolution branch of the former stage, and the number of channels is doubled; the feature maps of the parallel branches have different scales due to different resolutions, so they can focus on targets of different scales in the original image, and the backbone network fuses feature maps of different resolutions through upsampling and downsampling operations to realize information exchange between different levels and improve the multi-scale expression ability of the feature map.

[0058] When the detector is misdetected or missed due to occlusion in the pedestrian detection task, and only relying on the backbone network HRNet to extract pedestrian features cannot meet the accuracy requirement, the feature focusing module and the hierarchical fusion strategy are used to enhance the features of the occluded pedestrians by focusing and information fusion of the features.

[0059] In step S2, the feature focusing module performs fine processing on the input feature map from the dimensions of channel attention and spatial attention;

[0060] In terms of channel attention, the feature focusing module first performs global average AvgPool and maximum pooling MaxPool operations in the spatial dimension to condense two global description vectors to capture the global importance information of each channel in the entire feature map; then, a small fully connected neural network FC is used to learn and calculate the channel importance weight of each global description vector, the running results are added and normalized by the Sigmoid function, and a weight vector representing the relative importance of each channel is output Finally, the weight vector is multiplied by the input feature map f i to obtain a refined feature map F i 1 , which realizes the reweighting of feature responses; this process improves the sensitivity of the model to key feature channels, helps the network dynamically focus on information-rich channels, and enhances the representation ability and detection performance of the model;

[0061] In terms of spatial attention, the feature focusing module performs global average AvgPool and maximum pooling MaxPool in the channel dimension to focus on the potential information of the target region, then performs channel dimension splicing Concat, further compresses the spatial information through 7x7 convolution Conv, and uses the Sigmoid function for normalization to generate a single-channel spatial attention map Finally, the spatial attention map With a first-refinement feature map F i 1 Multiply and then weight the features again to obtain the secondary refined feature map F. i 2 By refining spatial processing, the sensitivity of the backbone network to key spatial locations is enhanced, enabling it to focus on target pedestrian areas and improving the network's target perception capabilities.

[0062] Step S2 uses a feature focusing module to dynamically adjust the importance of each channel and spatial position in the feature map, thus refining the feature information twice. This guides the neural network model to focus on the feature information of key regions and effectively suppresses noise interference. The specific formula is as follows:

[0063]

[0064] In the formula, f i Given multiple input feature maps of different scales, where (i = 1, 2, 3, 4), F i 1 f i Passage attention Weighted refined feature map, F i 2 For F i 1 Spatial attention The output feature map after reweighting.

[0065] like Figure 1 As shown, this example introduces a feature focusing module into GGFIE-Net, which not only optimizes the attention distribution of the feature map, but also enhances the model's ability to represent features of occluded pedestrians. This helps the model better understand the contextual information in the image, thereby achieving effective detection and recognition of occluded pedestrians.

[0066] In step S3, a hierarchical feature fusion mechanism is used to efficiently fuse the four feature maps with different resolutions output in step S2.

[0067] like Figure 1 As shown in the feature fusion section, this example first performs a preliminary fusion of three low-resolution feature maps containing rich contextual information, and then performs a comprehensive fusion with a high-resolution feature map containing more positional details to obtain a multi-scale feature map containing rich semantic information. Specifically: first, the three low-resolution feature maps containing rich contextual information are concatenated along the channel dimension to enhance the representation of deep semantic information; then, the first concatenated feature map is concatenated a second time with the high-resolution feature map containing more detailed positional information to achieve a comprehensive fusion of high-level semantic information and positional detail information; the specific formula is expressed as:

[0068]

[0069] T=Concat[F1 2 ,U 1 / 4 (t)] (6)

[0070] In the above formula, and The corresponding feature maps and the pixel ratio of the original input image are 1 / 4, 1 / 8, 1 / 16 and 1 / 32 times, respectively;

[0071] and indicate that the feature map and is upsampled to 1 / 8 scale using the bilinear interpolation algorithm, and Concat indicates that the channel dimension is spliced to form a preliminary fusion feature map Then U 1 / 4 (t) is upsampled to 1 / 4 scale using the bilinear interpolation algorithm, and finally spliced with the high-resolution feature map F1 2 in the channel dimension to obtain a comprehensive fusion output feature map

[0072] The specific method of step S4 is: adding a detection head after the spliced feature map to detect it; the feature map used for detection first passes through a 3x3 convolution kernel after entering the detection head, and then three 1x1 convolution kernels are attached in parallel to generate a pedestrian center point feature map, a pedestrian scale feature map and a center point offset feature map, respectively; then, according to the corresponding scale in the center point feature map and the scale feature map, the pedestrian candidate box of the input image is automatically generated; finally, the position of the pedestrian center point is fine-tuned through the center point offset prediction.

[0073] Therefore, the pedestrian detection can be represented as:

[0074] Dets=H(φ det )={cls(φ det ),reg(φ det )} (7)

[0075] Where H(.) represents the detection head, and φ det represents the feature map used for detection; cls(.) is the predicted classification confidence, and reg(.) is the predicted scale and center point offset of the pedestrian.

[0076] In step S4, focalloss is used to train the target function L c of the pedestrian center point position prediction in the detection head:

[0077]

[0078] where p ij and α ij are defined as equations (9) and (10) respectively:

[0079]

[0080] where y ij represents the center point of the position (i,j), p ij ∈[0,1] represents the probability that the predicted coordinate point belongs to the center, and β and γ are two hyperparameters, which are set as β = 4 and γ = 4. M ij is a Gaussian heat map, aiming to reduce the uncertainty of a large number of negative samples around the positive sample;

[0081] The scale prediction is a regression task using SmoothL1 loss, as shown in equation (11):

[0082]

[0083] where s k and represent the predicted value and the true value of the scale of the kth target;

[0084] The offset prediction is also a regression task using SmoothL1 loss, as shown in equation (12):

[0085]

[0086] where o k and represent the predicted value and the true value of the offset of the kth target;

[0087] The total loss is composed of the center loss, the scale loss, and the offset loss; it is represented as equation (13):

[0088] L = λ c L c + λ s L s + λ o L o (13)

[0089] where λ c , λ s , and λ o are set as 0.01, 1, and 0.1 respectively.

[0090] Embodiment:

[0091] To fully verify the robustness of the method of the present application in dealing with occluded pedestrians, the present example was tested on the challenging pedestrian dataset CityPersons.

[0092] CityPersons is a pedestrian detection dataset in traffic scenes, containing 5050 images (2975 for training, 500 for validation, and 1575 for testing). The validation subset in CityPersons is divided into four occlusion subsets according to the visibility of the target, in order to more fairly compare the effectiveness of the method for dealing with occluded pedestrians. Specifically, for the target with pixel value greater than 50, the visibility in the interval [0.65, 1] is divided into the reasonable occlusion subset (Reasonable), the visibility in the interval [0.65, 0.9] is divided into the partial occlusion subset (Partial), the visibility in the interval [0.9, 1] is divided into the bare occlusion subset (Bare), and the visibility in the interval [0, 0.35] is divided into the heavy occlusion subset (Heavy). The validation of the occlusion subset can be used as an important indicator to judge the performance of the detector in dealing with occlusion.

[0093] In order to more comprehensively evaluate the performance of the detector in pedestrian detection, the present example uses MR-FPPI evaluation curve. The missed detection rate (MR) refers to the ratio of the prediction of negative samples (FN) in all true positive samples (TP+FN). FPPI refers to the ratio of false positive samples (FP) in each image on average in N images to be detected. The average value of 9 log(MR) in FPPI ∈ [10 -2 ,10 0 ] is selected and the exponential operation is performed to obtain MR -2 . The smaller the MR -2 , the higher the performance of the detector. In this paper, MR -2 is used as an evaluation index for CityPersons.

[0094] The hardware device on which all the experiments of the present example are based is NVIDIA PCleA100 GPUs, and the programming language and deep learning framework on which it depends are Python3.8 and Pytorch1.11.0.

[0095] The algorithm of the present example uses HRNet as the backbone network and uses AdamW as the optimization algorithm. In addition, image flipping, random scaling and cropping are used for data enhancement strategy in the data loading stage. The detailed parameter settings on the CityPersons dataset are introduced in detail as follows.

[0096] The input image size of this example is set to 640x1280 pixels. The pre-trained HRNet on ImageNet is used as the backbone network to train the proposed network model in this example. During training, the batch size is set to 16 (4 images per GPU, using 4 GPUs in total), and the initial learning rate is set to lr=2x e -4 The model is trained for 150 epochs, and the learning rate is reduced to 1 / 2 (i.e., lr / 2) every 50 epochs.

[0097] To verify the effectiveness of the method in the traffic scene, the performance of the method in this example and some advanced research algorithms on the occlusion subset of the CityPersons dataset is compared as shown in Table 1. It can be observed that the performance of the method in this example on all subsets is very good, especially on the Heavy subset, which achieves the best performance. Even compared with some methods that use occlusion processing strategies, such as RepLoss, MB-CSP+BIA-NMS, OR-CNN, CaSe+OTP-NMS, CrowdDet+OTP-NMS, etc., the method proposed in this paper also shows considerable competitiveness. Compared with the recently proposed competitive model CaSe+OTP-NMS using an additional improved NMS strategy, the detector proposed in this paper achieves a performance gain of 4.0% MR -2 on the Heavy subset. In addition, to ensure fair comparison with the baseline model (CSP+HRNet), this section reproduces the baseline network under the same configuration, which achieves different degrees of performance gain on Reasonable, Bare, Partial, and Heavy, and achieves a performance gain of 4.3% MR -2 on the Heavy subset. Compared with the traditional center point and scale prediction-based pedestrian detector on the four occlusion subsets, the proposed method has a larger performance gain due to the enhancement of key area pedestrian feature information, which helps the detector effectively infer the occluded pedestrians by extracting context information. Finally, we test the method on the original image of 1024x2048, and the detection speed of the proposed method is 0.17s / img, which is comparable to other detectors. The above experiments fully demonstrate that the proposed method can effectively distinguish target and background information to handle pedestrians with different occlusion degrees.

[0098] The above is only an embodiment of the present application, and does not limit the technical scope of the present application in any way. Therefore, any minor modification, equivalent change and modification made according to the technical essence of the present application to the above embodiment are still within the scope of the technical solution of the present application.

[0099] Table 1. Results of comparative experiments on CityPersons dataset

[0100]

[0101]

Claims

1. An occluded pedestrian detection method based on feature focusing and information fusion, characterized in that: The method first uses HRNet, a high-resolution convolutional neural network, as the backbone to extract features from the input image, ensuring that high-resolution features are preserved while extracting image feature information. Secondly, a feature focusing module dynamically adjusts the importance of each channel and spatial position of the feature map to enhance key region features and suppress background noise, allowing the backbone network to focus on the target region. Then, a hierarchical feature fusion mechanism efficiently fuses high-level semantic information and location detail information. Finally, the fused feature map is passed through a detection head to generate a center point feature map, a scale feature map, and a center point offset feature map. The feature map is then interpreted into detection results, outputting a pedestrian region with bounding boxes. The method includes the following steps; Step S1: Input image and perform feature extraction; Step S2: Design a feature focusing module to focus and enhance the features extracted by the backbone network; Step S3: Design a hierarchical feature fusion mechanism to efficiently fuse multiple feature maps of different resolutions output in step S2; Step S4: Use the detection head to interpret the feature map into detection results; In step S2, the feature focusing module refines the input feature map from two dimensions: channel attention and spatial attention. Regarding channel attention, the feature focusing module first performs global averaging (AvgPool) and max pooling (MaxPool) operations in the spatial dimension to condense two global description vectors, thereby capturing the global importance information of each channel on the entire feature map. Subsequently, a small fully connected neural network (FC) is used to learn and calculate channel importance weights for the two global description vectors respectively. The results are summed and normalized using the Sigmoid function to output a weight vector representing the relative importance of each channel. Finally, the weight vector The input feature map f is subjected to multiplication. i A refined feature map F is obtained. i 1 This enables the reweighting of characteristic responses; This process enhances the model's sensitivity to key feature channels and helps the network dynamically focus on information-rich channels, thereby improving the model's representation ability and detection performance. In terms of spatial attention, the feature focusing module performs global averaging (AvgPool) and max pooling (MaxPool) in the channel dimension to capture the latent information of the target region. Then, it performs concatenation in the channel dimension, further compresses spatial information through 7×7 convolution (Conv), and uses the Sigmoid function for normalization to generate a single-channel spatial attention map. Finally, the spatial attention map With a first-refinement feature map F i 1 Multiply and then weight the features again to obtain the secondary refined feature map F. i 2 By refining spatial processing, the sensitivity of the backbone network to key spatial locations is enhanced, enabling it to focus on target pedestrian areas and improving the network's target perception capabilities.

2. The occlusion pedestrian detection method based on feature focusing and information fusion according to claim 1, characterized in that: In step S1, the convolutional neural network HRNet is used as the backbone network to process the input image and output four feature maps f1, f2, f3 and f4 with different resolutions, which are 1 / 4, 1 / 8, 1 / 16 and 1 / 32 times the resolution of the input image, respectively.

3. The occlusion pedestrian detection method based on feature focusing and information fusion according to claim 2, characterized in that: The backbone network HRNet adopts a parallel architecture, consisting of four branches with different resolutions. When extracting image features, the initial stage uses a high-resolution sub-network for feature extraction, and then gradually adds parallel sub-networks from high resolution to low resolution. Specifically, in its four stages, the first stage contains only one high-resolution branch. From the second stage to the fourth stage, each stage adds a parallel branch. The input of the next stage comes from the output of each branch in the previous stage, and the resolution of the newly added branch is half that of the lowest resolution branch in the previous stage, and the number of channels doubles. The feature maps of the parallel branches have scale due to their different resolutions, enabling them to focus on targets at different scales in the original image. The backbone network fuses feature maps of different resolutions through upsampling and downsampling operations, enabling information exchange between different layers and thus improving the multi-scale expressive power of the feature maps.

4. The occlusion pedestrian detection method based on feature focusing and information fusion according to claim 2, characterized in that: In pedestrian detection tasks, if occlusion causes false detections or missed detections by the detector, and relying solely on the backbone network HRNet to extract pedestrian features is no longer sufficient to meet the accuracy requirements, a feature focusing module and a hierarchical fusion strategy are used to enhance the features of occluded pedestrians by focusing on features and fusing information.

5. The occlusion pedestrian detection method based on feature focusing and information fusion according to claim 2, characterized in that: Step S2 uses a feature focusing module to dynamically adjust the importance of each channel and spatial position in the feature map, thus refining the feature information twice. This guides the neural network model to focus on the feature information of key regions and effectively suppresses noise interference. The specific formula is as follows: In the formula, f i Given multiple input feature maps of different scales, where (i = 1, 2, 3, 4), F i 1 f i Passage attention Weighted refined feature map, F i 2 For F i 1 Spatial attention The output feature map after reweighting.

6. The occlusion pedestrian detection method based on feature focusing and information fusion according to claim 2, characterized in that: In step S3, a hierarchical feature fusion mechanism is used to efficiently fuse the four feature maps of different resolutions output in step S2. Specifically, the three low-resolution feature maps containing rich contextual information are first concatenated along the channel dimension to enhance the representation of deep semantic information. Then, the first concatenated feature map is concatenated with the high-resolution feature map containing more detailed positional information to achieve a comprehensive fusion of high-level semantic information and positional detail information. The specific formula is as follows: T=Concat[F1 2 ,U 1 / 4 (t)] (6) In the above formula, and The corresponding feature maps have pixel ratios of 1 / 4, 1 / 8, 1 / 16, and 1 / 32 times that of the original input image, respectively. and This indicates that the feature map is processed using a bilinear interpolation algorithm. and Upsampling to a scale of 1 / 8, Concat represents concatenation along the channel dimension to form a preliminary fused feature map. Then U 1 / 4 (t) The feature map t is upsampled to a 1 / 4 scale using a bilinear interpolation algorithm, and finally compared with the high-resolution feature map F1. 2 A fully fused output feature map is obtained by stitching together the features along the channel dimension.

7. The occlusion pedestrian detection method based on feature focusing and information fusion according to claim 2, characterized in that: The specific method of step S4 is as follows: A detection head is added after the concatenated feature map to perform detection; the feature map used for detection first passes through a 3×3 convolution kernel, and then three parallel 1×1 convolution kernels are added to generate a pedestrian center point feature map, a pedestrian scale feature map, and a center point offset feature map, respectively; then, pedestrian candidate boxes of the input image are automatically generated based on the corresponding scales in the center point feature map and the scale feature map; finally, the position of the pedestrian center point is fine-tuned by predicting the center point offset; thus, pedestrian detection is represented as: Dets=H(φ det )={cls(φ det ),reg(φ det )} (7) Where H(.) represents the detection head, φ det The feature map used for detection is represented; cls(.) is the predicted classification confidence, and reg(.) is the predicted pedestrian scale and center point offset.

8. The occlusion pedestrian detection method based on feature focusing and information fusion according to claim 7, characterized in that: In step S4, focal loss is used to train the objective function L for predicting the position of the pedestrian center point in the head detection. c : Where p ij and α ij They are defined as formulas (9) and (10) respectively: In the formula y ij p represents the point located at the center of position (i,j). ij ∈[0,1] represents the probability that the predicted coordinate point belongs to the center, and β and γ are two hyperparameters, which are set to β=4 and γ=4; M ij The Gaussian heatmap is designed to reduce the uncertainty caused by a large number of negative samples surrounding positive samples. Scale prediction is a regression task using SmoothL1 loss, as shown in Equation (11): In the formula s k and This represents the predicted and actual values ​​at the k-th target scale; Offset prediction is also a regression task using SmoothL1 loss, as shown in Equation (12): In the formula o k and This represents the predicted and actual values ​​of the offset of the k-th target. The total loss consists of center loss, scale loss, and offset loss; expressed as formula (13): L=λ c L c +λ s L s +λ o L o (13) In the formula λ c , λ s and λ o Set them to 0.01, 1 and 0.1 respectively.

Images