Anchor-free Multispectral Pedestrian Detection Based On Attention Mechanism

Pedestrian detection,as an important branch of object detection,has been an important research topic in the field of computer vision.And it plays an important role in practical applications such as unmanned driving and intelligent surveillance.There are factors affecting pedestrian detection performance in practical applications such as illumination insufficiency and bad weather,and it is hard to solve these problems with traditional singlemodal visible images.In contrast,infrared images are less influenced by illumination.Therefore,it is possible to combine visible and infrared images for multispectral pedestrian detection to compensate for the lack of single-modal features.In this thesis,we propose a multispectral pedestrian detection method based on anchor-free detection,which improves the model inference speed by using the anchor-free detection framework.Meanwhile,we propose to refine the features by using the attention mechanism to improve the quality of feature fusion.The experimental results show that the method in this thesis has good performance in both detection accuracy and detection speed.The research in this thesis includes two main aspects as follows:(1)To address the problems of low detection accuracy and slow detection speed in multispectral pedestrian detection,an anchor-free multispectral pedestrian detection method is proposed.Firstly,the Differential Modality Aware Fusion module is used to solve the feature blind fusion problem.Secondly,we propose to mix local and global attention mechanisms to enhance the quality of feature fusion while capturing the correlation between local and global features.Finally,a lightweight backbone network and an anchor-free detection framework are used to reduce the model complexity and improve the detection efficiency.The experimental results on KAIST,FLIR,and CVC-14 datasets show that the performance of the method in this thesis is significantly improved compared with the current state-of-the-art methods.(2)To address the problems of low quality of multispectral feature fusion and high miss rate of small-scale pedestrian detection,a multispectral pedestrian detection method based on the Scale-aware Permutated Attention module and Adjacent-branch Feature Aggregation module is proposed in this thesis.Firstly,four different attention sub-modules are used to combine into a Scale-aware Permutated Attention module to enhance the feature mapping of different scales in the Feature Pyramid Networks and improve the fusion quality of visible and infrared features adaptively by emphasizing important information or suppressing unimportant information.Secondly,the Adjacent-branch Feature Aggregation module is used to aggregate the features of adjacent layers so that the features between different scales are fully aggregated,which enhances the robustness of the model to scale variance and improves the small-scale pedestrian detection accuracy.Finally,the detection speed is balanced using the anchor-free detection framework.Experimental results on KAIST and FLIR datasets show that the method in this thesis outperforms current state-ofthe-art methods.