Research On Pedestrian Search And Recognition Algorithms In Complex Scenes

The research of pedestrian search and recognition algorithms is of great significance to the maintenance of public safety.The pedestrian search and recognition system is composed of three modules: target pedestrian area positioning,candidate pedestrian detection in the area,and candidate pedestrian identification.The specific process is: first locate the target area through cross-domain retrieval and call up the video and picture information in the camera of the area,and perform pedestrian detection based on this information to obtain candidate pedestrians.Then input the candidate pedestrian image into the existing pedestrian library for re-recognition,and obtain a collection of pictures belonging to the same pedestrian.Finally,these pedestrian picture sets are input into the gait recognition module to identify the identity of the target pedestrian.The realization of the main phased functions in the system is based on scene-based cross-domain retrieval and positioning,dense pedestrian detection,pedestrian re-identification and gait recognition algorithms.In the algorithm landing stage,it is often faced with the test of complex scenarios.The cross-domain retrieval and positioning algorithm of the scene will cause the retrieval and positioning effect to decrease due to the difference of the image perspective of the two domains.When there is a high-density pedestrian scene,pedestrian detection will have the problems of prediction frame offset and error by non-maximum suppression algorithms.The recognition effect of pedestrian re-recognition is largely limited by complex situations such as the posture change and occlusion of the human body representation.Although gait recognition is not affected by characterization,the recognition effect declines sharply under the influence of complex factors such as pedestrians wearing coats.In summary,current pedestrian search and recognition algorithms cannot achieve good results in complex scenes.In order to solve the problem of large changes in perspective in scene cross-domain retrieval,this paper proposes a dynamic hierarchical feature selection algorithm(Dynamic Hierarchical Feature Selection,DHFS)to improve the current scene cross-domain retrieval method.This algorithm can fully extract feature maps.These hierarchical features in the network make the features of the network output more discriminative,thereby improving the accuracy of cross-domain retrieval.In order to enable the dynamic hierarchical feature selection algorithm to be flexibly embedded in the network,this paper also proposes the corresponding Adaptive Residual Structure(ARS).For the two tasks of satellite image retrieval drone image and drone image retrieval satellite image,DHFS was experimented on the dataset University-1652.The experimental results showed that Recall@1 reached 86.88% and 76.67% respectively,exceeding the current best effect.In order to solve the problem of pedestrian occlusion in pedestrian detection in dense scenes,this paper proposes a detachable Chain Match Detection(CMD).The algorithm can associate the pedestrian's head with the whole body through the visible part,where the head that is difficult to be occluded plays a role in promoting the detection of the whole body,and when inferring,the unneeded detectors(heads)can be determined according to specific tasks.CMD conducted experiments on the Crowdhuman dataset,and the results showed that the average precision(Average Precision,AP),the loss rate(Miss Rate,MR)and the Jaccard Index(Jaccard Index,JI)all exceeded the current best results.Reaching91.57%,41.08% and 83.07% respectively.In order to solve the problem of pedestrian posture and occlusion in pedestrian rerecognition in complex scenes,this paper proposes a novel cross attention mechanism algorithm(Cross Attention Module,CAM).This algorithm can coordinately locate the most discriminative local human body features based on the human body features observed from different perspectives.At the same time,in order to fully explore the relationship between the local features,a Part Triple Metric Loss(PTML)corresponding to the local features is proposed.The proposed algorithm has reached the current best results on the data sets Market-1501,Duke MTMC-re ID and CUHK03,and the Rank@1 indicators are95.03%,91.1% and 73.23%,respectively.In order to solve the problem that the gait recognition algorithm cannot recognize normally under complex conditions,this paper proposes a Local Feature Flow Regulation(LFFR).Because the frequency and amplitude of the movement of different horizontal parts of pedestrians are quite different when walking,the LFFR algorithm can establish a spatiotemporal relationship among the local features between frames,and adjust the weight of the local features based on this relationship.For gait sequences with larger motion amplitudes,the local features of the horizontal parts will be enhanced and input to the next stage,which makes the final features more discriminative.Experiments show that the effect of the LFFR algorithm in complex scenes such as pedestrians wearing coats greatly surpasses the current best gait recognition network,and the Rank@1 on the dataset CASIA-B reaches 74%.