Research On Pedestrian Gait Recognition Based On Deep Learning

In recent years,the field of computer vision has witnessed unprecedented development.In the meantime,gait-based identification technology has gradually attracted attention.Gait recognition refers to the identification of individual identity by walking.Gait features can achieve long-distance identification and non-contact acquisition,which has a large application potential in the fields of public security and smart city construction.This thesis mainly studies the gait recognition algorithm focusing on deep learning:(1)A study on gait contour map extraction method based on deep learning is carried out.Aiming at the problem that traditional gait contour extraction methods are sensitive to light and background changes and are easily affected by noise,this thesis uses Mask R-CNN,a deep learning detection segmentation model,to extract gait contour.Pedestrian segmentation data set is utilized for training,realizing the real-time and effectively segmentation of the human gait contour map.(2)A study of gait recognition algorithms based on deep learning is conducted.Aiming at the problems of low flexibility and ignoring the timing information in the current gait template method,an improved Gait Set is proposed for gait recognition.On the basis of Gait Set,the vertical-horizontal pyramid pooling module is proposed.Softmax loss function is introduced for joint training to overcome the shortcoming that the triplet loss function does not consider the compactness of inclass features.The improved algorithm has a certain improvement in the accuracy on gait recognition Dataset B,especially under the condition of walking with a coat.(3)Development of gait recognition system.Both the gait contour extraction method and gait recognition algorithm based on deep learning are integrated into a complete gait recognition algorithm.With the cooperation of gait contour segmentation and gait recognition algorithm,the system can recognize the gait of the object in the surveillance video and write the recognition results into the system database.