Research On Road Element Detection Technology For Unmanned Driving

Autonomous driving is a technology for automatically operating an unmanned ground vehicle(UGV)through sensing,positioning,computing,and control devices that equipped on the vehicle.It involves comprehensive application of artificial intelligence,computer vision,and automatic control technologies.Among them,road element detection based on computer vision is the core technology for unmanned equipment to sense the road environment and make control decisions,which is related to the safety and efficiency of vehicle driving.In order to realize the fast and accurate road element detection,this thesis deeply analyzes the visual characteristics of road elements,and proposes a series of road element detection methods based on deep learning,which greatly improves the accuracy of road element detection in autonomous driving scenarios.This thesis focuses on three main types of road elements:traffic signs,pedestrians,and road surfaces.Because the three road elements have unique visual characteristics,the detection algorithm also faces different challenges:1)Traffic signs usually have small visible areas,different traffic signs have different shapes,sizes,colors;2)the distance between pedestrians and the vehicle,and the diverse behaviors of pedestrians make extreme scale variation and occlusion;3)road pavements have different visual characteristics due to different materials,time of use,road grades,etc.,which greatly affects the generalization ability of the detection model.To overcome the above challenges,this thesis presents a series of effective detection models for traffic signs,pedestrians and pavements:firstly,we propose a local context network to learn a multi-scale local feature for small targets,which achieves improvement for traffic signs detection in suburban and urban roads,respectively.Then,we design a pedestrian detection model based on the feature pyramid network,which significantly improved the accuracy on the WIDER pedestrian detection dataset.Finally,we introduce pixel-level features into the detection network,which significantly improves the generalization capacity of model for road surface detection.In summary,this thesis analyzes the visual characteristics of road signs such as traffic signs,pedestrians and road pavements in the autonomous driving scene,and presents effective detection methods for different road elements.Through extensive experiments on public datasets,i.e.,Tsinghua-Tencent 100K,BDCI16-TSDAD2,and WIDER,and Bupt road dataset collected from real road scenes,we validate the performance of the proposed road element detection methods.