Research On Multispectral Pedestrian Detection Based On Anchor-Free Method

Pedestrian detection is an important research direction in the field of computer vision.It has been widely used in video surveillance,intelligent transportation,automatic driving and other fields.Traditional pedestrian detection methods use visible image as input,and its performance will decline sharply in the case of low illumination or overexposure.The pedestrian detection method based on multispectral images can make full use of the information complementarity of multi-source images,and can better adapt to all-weather scenes.In recent years,the anchor-free detection method shows good robustness,and can achieve the balance between detection accuracy and running speed.It has good application value in actual traffic scenes.Based on this,this thesis realizes the anchor-free pedestrian detection methods suitable for multispectral images,studies the core technologies such as anchor-free detection head network,feature extraction and fusion,and puts forward corresponding solutions to some problems.The main work is as follows:1.Aiming at the shortcomings of high computational complexity and slow reasoning speed of anchor-based pedestrian detection method,an anchor-free multispectral pedestrian detection method based on Modality Balance Network(MBNet)is proposed.Firstly,the method adopts the illumination sensing module to provide illumination weights for subsequent training according to the illumination conditions of the image,so as to make the model adapt to different illumination environments.Secondly,in order to solve the problem of modal imbalance caused by multispectral data sources,the Differential Modality Aware Fusion(DMAF)module is used to enhance the information,realize the feature compensation between multimodal information,and reduce the imbalance of modal information,so as to improve the robustness of the method.Finally,the Full Convolutional One-Stage Object(FCOS)detection network is used to realize pixel by pixel detection and complete the anchor-freedetection task.The experimental results on KAIST dataset show that this method can achieve the balance between accuracy and processing speed.2.Aiming at the problems of low image resolution and pedestrian occlusion in traffic scene,an anchor-free multispectral pedestrian detection method based on feature selection is proposed.Firstly,the pedestrian detection model based on traditional feature fusion method is constructed.Secondly,the feature selection strategy is adopted and three pedestrian detection models based on feature selection are constructed.Finally,the attention mechanism is introduced into the network to enhance the pedestrian target feature information.The experimental results show that the feature selection strategy is better than the traditional feature fusion method,the detector performance is robust,and is suitable for real-time scene detection.