Research On Algorithm Of Pedestrian Pose Estimation And Re-recognition Based On Deep Learning

With the rapid development of artificial intelligence technology and the great potential of practical application,how to realize the pedestrian pose estimation and pedestrian reidentification in surveillance video,timely discover and deal with abnormal dangerous behaviors of pedestrians and realize search and tracking of object people,thereby improving public The safety early warning capability of the Public Places has become one of the hot research topics in industry and academia.The traditional and existing mainstream deep learning-based pedestrian pose estimation algorithms cannot meet the real-time and accuracy requirements of task processing when applied to actual surveillance video occasions.In addition,the pedestrian re-identification algorithm pays more attention to the study of singleframe images and global features of images.When applied to actual surveillance video,it cannot effectively use the information contained in the video sequence and process the background noise caused by pedestrian movement.In response to these problems,this paper focuses on the Pedestrian Pose Estimation and Re-identification algorithm based on deep learning.For the research of pedestrian attitude estimation,this thesis proposes a multi-task based pedestrian attitude estimation algorithm.Firstly,the lightweight YOLOv3-based pedestrian object detection network and the feature pyramid-based pedestrian key detection network are merged in parallel to one end-to-end network.In the training and prediction network,multi-task simultaneous detection is realized,which improves the running speed of the algorithm.Then,using the detected bounding box and key points of the pedestrian target,the human body posture is learned through the network structure of the pose residual,and the ambiguity problem caused when the key points are classified to the individual instance due to the overlap of the multi-person bounding box is solved,and the key point can be clustering grouped accurately and improved the accuracy of pedestrian pose estimation.The experimental results show that compared with the pose estimation algorithm based on partial affinity Fields,the multi-task based pose estimation algorithm proposed in this paper improves the running speed of Frames Per Second(FPS)by about 2 times,reaching 20 FPS,with real-time processing effect,and the average accuracy rate increased by about 2% to 67.2% on the MSCOCO test set.For the research of pedestrian re-identification,this thesis proposes a pedestrian reidentification algorithm based on multi-dimensional local feature aggregation.Firstly,based on video image sequence,the pedestrian re-identification algorithm is used to aggregate multi-frame pedestrian image features.Using our proposed multi-task pedestrian pose estimation network built a multi-task region proposal network.According to the detected pedestrian pose and bounding box,the pedestrian local area division proposals were obtained,which reduced the background noise interference and solved the pedestrian image alignment problem.Next,the image local feature generation network was built.It is recommended to cut the pedestrian image and extract the local features of the image by using the convolutional neural network,and use the image quality assessment network to evaluate the local quality of the pedestrian image.Finally,the feature aggregation unit is constructed,and the image sequence is multi-dimensional local feature aggregation according to the local image quality score.The features of different local regions are aggregated according to the quality scores,and the corresponding local region features complement each other,which reduces the influence of noise regions such as blur or occlusion and obtains more representative pedestrian video-level discriminant features.The experimental results show that the multi-dimensional local feature aggregation proposed in this paper improves the pedestrian re-identification accuracy,especially in the environment of multi-noise interference such as blur or occlusion.Compared with the QAN(Quality Aware Network)network using the global features of the image,the PRID 2011 data set and the i LIDS-VID data set.Rank-1 increased by 0.9% and 8%,respectively,and rank-5 increased by 0.3% and 4.4%,respectively.The multi-task based pedestrian pose estimation and multi-dimensional local feature aggregation based pedestrian re-identification method which proposed in this thesis can be used in intelligent video surveillance scenarios in public places.