Anchor-free Pedestrian Detection And Its Implementation

With the development of visual object detection,its subtask pedestrian detection has attracted the attention of scholars.Pedestrian detection has important applications in the intelligent industry,which has extremely high requirements for the accuracy and speed of the algorithm.This task is closely related to life,so there has been no effective way to solve the high miss rate caused by the complicated background and the problems of dense and overlapping objects.Therefore,this thesis focuses on the problems of object crowding and occlusion in pedestrian detection,proposes two pedestrian detection algorithms based on anchor-free network and designs a pedestrian detection system.The specific work is as follows:(1)An adaptive normalized feature aggregation network based on anchor-free network is proposed.To analyze the normalization method,this thesis proposes an adaptive normalization method.Specifically,this module enables different convolutional layers of the network to automatically select the most appropriate normalization method with a very small cost.To solve the problem of dense and overlapping objects,the network introduces feature aggregation module to aggregate the feature information of the same object to shorten the distance between the predicted object and the ground truth.The effectiveness of the network is proved by the experimental results on the public pedestrian detection dataset.(2)A cascaded cross-layer fusion network based on anchor-free network is proposed.To analyze the method of object detection neck,this thesis proposes a cascaded cross-layer fusion module.Specifically,this module denoises shallow features by cascading and reusing deep semantic information to retain more features.To solve the problem of object crowding and occlusion,the network introduces global smooth map to make the center graph have better location capabilities.The proposed network achieves the robust performance on the Caltech and City Persons datasets.(3)A pedestrian detection system based on anchor-free network is designed.The system has guided interactions,allowing users to quickly use it.Meanwhile,the system integrates the above two network models.The users not only needs to input one image to get the detection result,but also can choose whether to persist according to their needs.