Multi-Features Fusion Based Pedestrian Detection Method

Pedestrian detection,as an important branch of object detection which is a hot issue in the field of computer vision and artificial intelligence,has widespread application value in advanced driving assistance systems,intelligent video surveillance,intelligent transportation,etc.As a deformable target,pedestrian has the characteristics of both rigid object and soft object,and its detection accuracy will easily be changed owning to the changes of posture,appearance,movement and perspective.Single feature cannot fully depict pedestrian,so the accuracy of pedestrian detection algorithm which is based on single feature is not good enough.In this thesis,we propose a pedestrian detection algorithm based on multi-feature fusion through feature space fusion and multi-channel feature fusion,thus to improve the performance of pedestrian detection accuracy and speed.The experimental results show that the proposed algorithm has better detection performance compared with a single feature.The work of this thesis mainly focuses on the following aspects:(1)When HOG-ULBP feature which is aimed at gradient and texture depiction is used in the pedestrian detection,the fusion feature dimension we get is too high with too many redundant information,which affects the detection speed and the detection accuracy.To solve this problem,we propose a new method to reduce the dimension of HOG feature by using PCA method,and then form a new fusion feature by cascading the new HOG feature and ULBP feature,thus to effectively improves the performance of the pedestrian detection algorithm.(2)Aiming at the problem that the HOG-ULBP pedestrian detection algorithm is difficult to meet the real-time requirement,this thesis proposes a weight-based cascade classifier fusion algorithm.We propose to train the weak classifier based on HOG and the weak classifier based on ULBP respectively,and then combine two weak classifiers to form a strong classifier through weight cascade.The pedestrian area to be verified is output by using the ULBP weak classifier of the first stage,and the pedestrian area is verified by using the strong classifier to output the final pedestrian target position,thus to effectively improves the speed of the pedestrian detection algorithm.(3)When it comes to problem that the detection window fusion algorithm based on non-maximum suppression(NMS)can not eliminate erroneous detection window,this thesis proposed window fusion algorithm based on NMS to select candidate window,and then make further erroneous detection window elimination by adjacent window detection and the comparison of confidence to reduce the false detection rate.(4)Apply multi-channel feature on pedestrian detection through the way of channel feature fusion to improve pedestrian detection speed.Since video transmission uses YUV format encoding the thesis choose YUV color channel as color channel feature instead of the traditional LUV color channel feature,and integrate with the gradient histogram feature and the gradient magnitude channel feature,thus to get a new channel feature.In order to solve the problem of slow detection,this thesis takes the methods of integral histogram to quickly compute channel feature.With multi-scale feature estimation,this thesis could quickly compute the multi-scale channel feature of images to improve the detect speed.(5)When it comes to the characteristics of multi-channel features,this thesis proposes to adapt different normalization scheme according to different characteristics of channel feature.For example,the gradient histogram feature is normalized by the gradient magnitude,and the color channel feature and the gradient magnitude channel feature are normalized using the pixel region relationship,thus to improve the performance of pedestrian detection.