Multi-scale Pedestrian Detection Based On Deep Convolutional Feature Fusion

As a special task distinguished from general object detection,pedestrian detec-tion has great academic value and broad application prospect in smart surveillance,automatic driving and intelligent robot.Despite the significant improvements in re-cent years,there are still many challenges in pedestrian detection.In many application scenarios,the scale distribution of pedestrians is very wide,and small-size and medium-size pedestrians account for a large proportion.However,the existing pedestrian detec-tion approaches perform poorly in small-size and medium-size pedestrian detection.In this thesis,we carry out research on multi-scale pedestrian detection and aim to improve the detection accuracies of small-size and medium-size pedestrians.Inspired by the classic general object detection framework,we propose an end-to-end multi-scale pedestrian detection approach.First of all,we put forward the region proposal network for multi-scale pedestrian detection according to the characteristics of pedestrian.The proposed network can generate high-quality multi-scale pedestrian candidate regions via multi-scale pedestrian reference boxes and small-step "anchor".Secondly,we propose a region-of-interest pooling method and a multi-level deep con-volutional feature fusion method in response to the lack of small-size and medium-size pedestrian feature information,which can combine low-level features and high-level semantic features to enrich the representation of pedestrian.Then,considering the com-plex environment and many hard examples,the focal loss function is introduced to carry out hard example mining and further to improve the detection accuracy.Finally,we train the multi-scale pedestrian detection model through the joint training method to-gether with some common training tricks.We conduct extensive experiments on Caltech dataset,INRIA dataset and ETH dataset to verify the effectiveness of the proposed approach.The multi-scale pedestrian detection approach achieves very competitive results and has a great advantage in small-size and medium-size pedestrian detection.At the same time,the work of this thesis also has a certain reference significance for some other special detection task,such as vehicle detection.