Multi-feature Learning Approach For Pedestrian Detection

Human vision system(HVS)is made up of visible light receiving organs and visual nervous systems,which is the most important sense receptor to construct the perception of surrounding world.Computer vision hopes to accomplish the work of vision system through optical imaging equipment and digital processing chip,including image processing,object detection,object tracking and image retrieval.Pedestrian detection,as the core technology of security monitoring,intelligent transportation,human-computer interaction,is a hot research direction that last over ten years.The main study of this thesis is the pedestrian detection method oriented to vision in the common scene.The effective features of the pedestrian area are obtained from the video frames obtained from the image acquisition device,and then machine learning algorithm is applied to identify the pedestrian area,so as to achieve the purpose of machine automatic identification of pedestrians.It is an important part of automatic driving and human-computer interaction.The real-time,accurate and universal pedestrian detection algorithm is a hot research direction in the field of computer vision.The concrete research contents are as follows:(1)The object detection method based on vision is studied.This method can be summed up as the combination of image feature extraction and machine learning algorithm.By studying the representative method of universal object detection,we have understood the inherent logic of visual identification method,and conducted in-depth research on object location and effective image features.The classical model of pedestrian detection,the statistical learning algorithm and the latest method of object detection based on deep learning are studied.Understanding and mastering the previous research results,noticing the existing problems of traditional methods,we can take advantages and avoid weaknesses of traditional methods in the research process,and make reasonable improvements.(2)A pedestrian detection method based on multi-feature learning is proposed for visual pedestrian detection in the common scene.In the object location part,the visual saliency method is used to narrow the search area,and then the Edgebox method based on the object edge is used to generate the proposal box,which improves the quality of the classifier input.The pedestrian recognition part adopts the joint training based on edge features and learning features.Finally,the inter-frame information fusion is proposed to correct the classification confidence value,so as to reduce the miss rate of the detector.The subjective and objective tests of the algorithm show that the proposed method perform better.(3)The method of deep learning detection based on regression is studied and introduced into the research of pedestrian detection.After analysis,this thesis holds that the main difficulties of pedestrian detection in the common scene are large scale and hard negative example.Therefore,we have made two improvements in the training of the deep learning model: join the multi-scale proposal box and use the focal loss to retrain the model.Finally,the test of the algorithm shows that the proposed approach presents competitive accuracy and efficiency.