Research On Pedestrian Detection Based On Deep Learning

Pedestrian detection is a challenging task in computer vision,which aims at accurately identifying and localizing all pedestrian instances in a given image.It has been widely used in many applications such as intelligent surveillance,autonomous driving and robotics.With the development of deep learning technology,pedestrian detection has witnessed significant progress in recent years.However,when considering the diverse requirements of actual applications,there is still much room for improvement.This thesis focuses on occlusion handling and fully end-to-end detection that discards heuristic Non-Maximum Suppression(NMS)post-processing.Meanwhile,this thesis also aims to achieve a balance among the accuracy,efficiency and easy deployment of algorithms.The main contributions of this thesis are listed as follows:(1)To tackle occluded pedestrian detection,this thesis proposes a Progressive Refinement Network(PRNet),which progressively detects the full-body bounding boxes of pedestrian instances by three phases:visible-part estimation,calibrating detected visible-part boxes to a full-body template derived from occlusion statistics,and full-body bounding boxes refinement.In addition,an occlusion loss and a Receptive Field Backfeed module are proposed to facilitate training.To further improve the model’s overall performance of occlusion handling,this thesis proposes PRNet++,which is an improved detector with a dual-stream architecture,where an easy-branch and a hard-branch are designed to learn complementary representations that are more robust to various occlusions.To further explore the occlusion handling issue under cross-domain situations,this thesis introduces the unsupervised domain adaptive occluded pedestrian detection task.To handle occlusion situations in unknown domains better,this thesis proposes a Dynamic Iterative adaptation strategy and a Multi-Experts adaptation strategy.Extensive supervised within-dataset experiments,crossdataset testing experiments and unsupervised domain adaptation experiments validate the effectiveness and superiority of the proposed methods.(2)To realize fully end-to-end pedestrian detection,this thesis proposes an Optimal Proposal Learning framework.It achieves the goal by two novel modules:a Coarse-to-Fine learning strategy for exploring the best classification decision-boundary via progressive boundary refinement,and a Completed Proposal Network for producing extra information compensation to help recall hard pedestrian samples.Extensive experiments demonstrate that the proposed end-to-end pedestrian detector can yield state-of-the-art performance.(3)To handle the requirements of extension tasks in actual applications,this thesis establishes a high-performance end-to-end detector.While achieving good performance on pedestrian detection,it can also tackle the detection tasks for objects of other required categories.In detail,an Enhanced Positive Sample Filter is proposed to filter out the single positive sample for each GroundTruth instance.This is realized by two components:a Dual-stream Feature Enhancement module that provides rich information clues,and a Disentangled Max Pooling Filter that enhances the feature discriminability in potential foreground regions.Equipped with EPSF,the proposed end-to-end detector shows great performances on both pedestrian detection and object detection tasks.In summary,this thesis mainly proposes to tackle occlusion issue and NMS-free problem based on the one-stage detection framework,which promotes the development of state-of-the-art pedestrian detectors and supplies references for actual applications.