Pedestrian Detection And Attitude Estimation Based On Deep Learning

With the increase of car ownership year by year,road safety has become an important issue.The loss of casualties caused by traffic accidents is very large every year.It is of great social significance to develop new technologies to reduce the loss caused by traffic accidents.From the perspective of pedestrian protection,this paper studies the task of pedestrian detection and human posture estimation based on monocular camera,which provides algorithm support for intelligent vehicle to understand the pedestrian status in the road.The specific research content of this paper is as follows.In view of the limited computing power of embedded system,based on the YOLOv3 algorithm,this paper proposes the PRUNE-YOLOv3 pedestrian detection algorithm,which has the same detection accuracy as YOLOv3 and fast detection speed.First of all,the YOLOv3 model is pre trained on the open source pedestrian detection data set,the pre training model is sparse trained on the self-made data set,the model is pruned at the channel level,and the unimportant channels are deleted.At this time,the model has been reduced to a certain extent,because the channel pruning will cause certain damage to the model,resulting in the degradation of the model performance,so it is necessary to The model is fine tuned to repair the detection accuracy of the model,and then the performance of the model needs to be evaluated.When the performance of the model is not up to the requirements of practical application,the pruning process is cyclic until the model is stable and meets the requirements of practical application.Finally,the YOLOv3 algorithm and the PRUNE-YOLOv3 algorithm are tested on the same dataset.Compared with the YOLOv3 model,the parameters of the proposed PRUNE-YOLOv3 model are reduced from 62.5M to 6.13 M,the volume of the model is reduced from 246.4MB to 25.6MB,the test time of a single image is reduced from15.7ms to 5.84 ms,but the AP(average precision)of the model is also reduced from0.7489 to 0.7245.Overall,the detection speed of the model has been improved by 2-3times,but the detection accuracy has slightly decreased.Based on the RMPE(regional multi person pose estimation)algorithm,this paper proposes a P-RMPE algorithm,which uses the PRUNE-YOLOv3 algorithm as the detector instead of the original SSD-512 algorithm.This algorithm takes the result of PRUNE-YOLOv3 detection as the input,designs a symmetrical space network to solve the problem of inaccurate pedestrian detection and positioning,then uses hourglass single person attitude estimation algorithm,finally uses the parameterized attitude non maximum suppression algorithm to deal with redundant attitude,and outputs the final attitude estimation result.Under the same experimental conditions,the performance of the two algorithms is compared.It can be seen from the experimental results that the detection accuracy of P-RMPE algorithm for each joint point of human body is slightly reduced,but the detection speed of the model has been greatly improved.