Study On Algorithms Of Person Search Based On Person Detection And Re-identification

With the rapid development of computer vision and artificial intelligence technology,person search based on person detection and reidentification has become a new hot-spot.Person search refers to the process of finding the person whose identity matches target person in a set of natural scene images.Specifically,it first detects the locations of every person appearing in an image,then it extracts features of each person and compares each of them with the target one.Therefore,person search can be seen as an extensive problem of person reidentification,which involves both person detection and reidentification.Nowadays,most of the reidentification algorithms only work on datasets that all person images are cropped manually.As a result,these person reidentification algorithms are not suitable for real-world applications.Besides,during person search,the precision of person detection is vital to the following reidentification task,which is sensitive to detection of error object,missed detection and bounding boxes of persons with low precision.Hence,how to realize the end-to-end person search method by combining person detection and reidentification is a challenging problem.This thesis probed into the integration of person detection and reidentification based on existing methods,and proposed some new methods to solve the problems appearing in these methods from the perspective of bounding box selection,the design of similarity function as well as the improvement of loss function.The main work and research results of thesis are as follows:1.Aiming at the precision of bounding boxes generated by existing person search algorithms during person detection,and the limitation of cosine distance when measuring feature similarities of different persons in person reidentification stage.This thesis proposed a new person search method by fusing the person detection and re-identification modules based on the modified Faster R-CNN.Firstly,it used an iterative bounding box regression network to promote the precision of bounding boxes.Then to enhance similarity learning ability,it used a modified metric learning method named MSLF which consists both cosine distance and Euclidean distance.Finally it added center loss to the whole loss function of network.Center loss boosts the network's ability by extracting discriminative features of different persons,and enables the network to achieve a better result for query person search.It performed simulation on a large scale benchmark dataset named CUHK-SYSU,the experimental results show that proposed method achieves 81.6% in CMC top-1,and 78.9% in m AP,which outperforms other paralleling methods about 0.4%?18.9% in CMC top-1 and 1.0%?23.2% in m AP.2.While extracting the features of different persons in person search network,some particular locations such as clothes,backpack and face are not stressed during feature extraction.Besides,the Softmax function has limitation when classifying different persons,because it can't minimize their intra class distance and maximize inter class distance.Therefore,this thesis used the attention mechanism and multi-loss function to train the person detection and reidentification modules separately,which improved the quality of feature vectors of candidates extracted from Faster R-CNN.Meanwhile,margin loss was proposed to further enhance the similarity learning ability of the network with center loss,enabling the network to achieve a more accurate result for target person search.Therefore this thesis proposed another person search method based on attention mechanism.Firstly,the algorithm utilized Faster R-CNN based on VGG16 to get the candidates.Then those candidates were sent into the Res Net50 containing attention mechanism to get the corresponding features,which finally were compared with the feature of query person.Using the large scale benchmark dataset named CUHK-SYSU to simulate the proposed method,the experiments results show that the proposed method achieve 83.5% in CMC top-1,and 80.4% in m AP,which outperforms other paralleled methods about 2.3%?20.8% in CMC top-1 and 2.5%?24.7% in m AP.