Study Of Pedestrian Detection Method Based On Deep Learning

In recent years,deep learning added a huge boost to the already rapidly developing field of artificial intelligence and computer vision.As an important task of artificial intelligence and computer vision,pedestrian detection plays an important role in driver assistance system and automatic driving.Focusing on pedestrian detection,this paper studies pedestrian detection methods based on deep learning.Most of the existing public pedestrian detection datasets are collected in Europe or the United States.There are few pedestrian detection databases based on Asian scenes.In order to increase the diversity of existing pedestrian detection databases,we build VIP pedestrian detection dataset.Many existing methods for pedestrian detection have achieved high accuracy on large-scale pedestrian detection,but they do not perform well on small-scale pedestrian detection.In this paper,we propose a novel deep small-scale sense network(termed as SSN)for smallscale pedestrian detection.With the special fusion architecture,our method improves the performance for small-scale pedestrian detection.The proposed architecture could generate some proposal regions which are more effective to detect small-scale pedestrians.Furthermore,we design a novel loss function based on cross entropy loss to increase the loss contribution from hard-to-detect small-scale pedestrians.In addition,a novel evaluation metric is introduced,which can measure the location precision of the pedestrian detection methods.Our method achieves good detection performance on Caltech pedestrian dataset and our VIP pedestrian dataset.Detecting occluded pedestrians is difficult for the existing methods.In this paper,we propose a novel judgement network for occluded pedestrian detection.The judgement network could judge the validity of the features after fusion,and select more effective features for occluded pedestrian detection.In addition,we construct a mask network to make features selected by the judgement network more robust.The experimental results verify the effectiveness of the proposed method.The existing pedestrian detection methods based on deep learning cannot satisfy the speed need in practical application.In this paper,we propose a separate convolution structure and step convolution structure to solve the speed problem for pedestrian detection methods based on deep learning.The separate convolution structure and step convolution structure could reduce a large number of network parameters in convolution extraction.Our method achieves superior detection performance on Caltech pedestrian dataset and our VIP pedestrian dataset.