| Pedestrian re-recognition refers to the technology of using computer vision technology to match pedestrians based on the video collected by a multi-camera network with non-overlapping fields of view.It is a key task in intelligent video surveillance,and is applicable to the technical fields such as security and human search in public places.At the current stage,the method of using deep learning has been widely used in pedestrian re-recognition.However,in the bus operation scene,the pedestrian body parts are blocked by the dense crowd,and the traditional pedestrian re-recognition method is no longer applicable.Based on the deep learning technology,this paper takes the head area of the passenger image as the effective part of the re-recognition to solve the problem of how to extract the video sequence of the boarding and alighting passengers from the bus surveillance video,and take the video sequence as the research object to realize the re-recognition of the boarding and alighting passengers.The work content consists of the following two aspects:(1)Research on multi-target tracking algorithm of public transit passenger head based on detection.Using the multi-target tracking technology based on detection,the passenger head video sequence for re-recognition is generated from the on-board surveillance video.This part is mainly divided into data set construction and multi-target tracking algorithm research.In terms of data set construction,two data sets suitable for detector and multi-target tracking algorithm experiments are sorted out.The data set used in the detector experiment consists of 12067 training set pictures and 2372 test set pictures;The data set used in the multi-target tracking experiment consists of 3246 training set pictures and 1286 test set pictures.In the research of multi-target tracking algorithm,YOLO is used in the detector part,and multiple matching method is used in the data association part.The Kalman filter is used to predict the position of the tracking track of the current frame in the next frame,and the IOU between the prediction frame and the actual detection frame is used as the similarity of the two matching,and then the Hungarian algorithm is used to complete the matching.In the experimental part,the detector model can obtain mAP＿50 value 0.998,mAP＿According to the experimental results of 5095 value 0.8212,the multi-target tracking algorithm can obtain the experimental results of MOTA value 92.4% and MOTP value0.128.(2)Research on the re-recognition algorithm of bus passenger head video sequence based on multi-granularity hypergraph.The multi-granularity hypergraph network can model the bus passenger head video sequence according to the space-time dependence in multiple granularity,and solve the problem of feature extraction of image information and time sequence information in the bus passenger head video sequence.This part of work is mainly divided into data set construction and research on the algorithm of bus passenger head video sequence recognition based on multi-granularity hypergraph.In terms of data set construction,the data set applicable to the research of bus passenger head recognition algorithm based on video sequence is sorted out,which is MARS data set format.The main content is the track information of bus passenger head video sequence,which is composed of 4862 training set pictures and 3811 test set pictures.In the aspect of algorithm research,the method of multi-granularity network is used to realize the feature extraction and granularity division of bus passenger head video sequence images,and the method of hypergraph neural network is used to realize the aggregation of grain-level features of video sequence images,and finally the re-recognition is completed according to the aggregated feature values.The re-recognition algorithm of bus passenger head video sequence based on multi-granularity hypergraph can obtain the experimental results of the mAP value of76.7% and the Rank-1 value of 81.4% on the test set. |
PDF Full Text Request