Research And Application Of Pedestrian Detection Method Based On Convolutional Head-shoulder Features

Pedestrian detection is one of the most popular researches in the field of computer vision in recent years.It has broad application prospects in the field of intelligent video surveillance,human-computer interaction systens and automatic driving.On the other hand,pedestrain detection is a basic research which is the foundation for follow-up works sunch as pedestrian tracking,person retrival and pedestrain behavior analysis.The results of pedestrain detection have a directly effect on the development of follow-up visual tasks.With the developments of Faster RCNN,SSD and YOLO,high performances have been achieved on Benchmarks such as Pascal VOC,COCO Detection,etc.It has become a trend to migrate these excellent algorithms to pedestrian detection.However this task itself has the following challenges compared to odinary object detection.Firstly,pedestrains vary widly in attitude,making it difficult to learn a unified feature description;Secondly,the scales of pedestrains change widly in one picture,the feature description cannot be adapted to scale changes;Thirdly,mutual Occlusions often occour in dense groups.Based on the in-depth analysis of the above problems,this paper proposes a pedestrian detection method based on the deep head-shoulder feateres on the base of Regionbased Fully convolutional neural network(R-FCN).This method mainly includes the following three contributions:(1)Aiming at the problem of pedestrian attitude,this paper utilizes the human head and shoulder parts which called Omega model as the representation model of the target pedestrain in the video surveillance.Compared with traditional human full-body model which widely used in academia,the human head shoulder parts model is more robust to posture changes,Besides,human head shoulder parts model can reduce the probability of being occluded.(2)Aiming at the problem of scale changes,this paper employs the location-sensitive pooling method in R-FCN and uses two paths with different reciptive field to detect largescale and small-scale human target separatelly,which makes the network be adapt to scale changes at the cost of introducing tinny computational overhead.(3)Aiming at the problem of partial occlusion,this paper proposes two improvements.On one hand,this paper proposes an improvement of the non-maximum supression algorithm witn bootstrap sampling strategy,it deal with pedestrian adjacent occlusion by introducing a kind of box similarity meature.On the other hand,from the perspective of changing the distribution of occlusion samples,this paper improves the adaptability of the network to pedestrian occlusion by generating more samples using generative adversarial network(ASDN).In this paper,the proposed method is compared with two representative deep learningbased object detection algorithms on three data sets.the proposed method achieves average precisions of 95.35%,96.27%,97.43% on Caltech dataset,DukeMTMC dataset and Bronze dataset respectively,which verifies the effectiveness of our proposed improvement in the actual surveillance scenarios.