Research On Pedestrian Detection With Occlusion Based On Deep Learning

Pedestrain detection refers to the accurate identification of pedestrians in pitures or videos through algorithms,which is a branch of object detection.Due to its wide application and great commercial value,it attracts a lot of attention from the government,universities and businesses.However,due to the non-rigid characteristics of pedestrians,the complex scenes,the change of shooting angle and other factors,pedestrian detection has a series of problems that need to be solved.Traditional pedestrian detection has always tried to find better ways to describe pedestrain features,but the achievement is not satisfactory.With the great success of deep learning,the pedestrian detection solution also transforms from traditional manual designed features to deep learning.The research in this paper is based on SSD which is current state-of-the-art object detection algorithm.This paper proposes two algorithms for pedestrian detection with occlusion.The main achievements of this dissertation are as follows:1.An algorithm of detecting pedestrians with occlusion is proposed based on Densebox Selection strategy and Repulsion Loss.The method consists of two parts,one is the improved Repulsion Loss and the other is the Densebox Selection strategy.Firstly,by analyzing the implementation details of Repulsion Loss designed for pedestrian detection with occlusion,it is found that there are some defects in the Repulsion Loss when it is used in SSD algorithm.The improved Repulsion Loss achieves good performance in the SSD algorithm by choosing the surrounding ground truth bounding box in different ways and setting overlapping thresholds.And then by analyzing the default box selection strategy and default box matching strategy of SSD,it can find that there are limitations when detecting pedestrians with occlusion.Therefore,a Densebox Selection strategy is designed for detecting pedestrians with occlusion,which improves the performance of SSD greatly.2.By analyzing the aspect ratio of pedestrians and the receptive field of convolutional features,a pedestrian detection algorithm based on multi-scale receptive field and Inception-ResNet is proposed.In order to reduce noise,this method generates feature maps through the Inception-ResNet module with multi-scale receptive field.Finally,we make some optimizations for the SSD network,making it suitable for pedestrian detection.This algorithm improves the performance of SSD greatly.