Study On Long Distance Small Objects Detection Algorithm On Structured Road Based On Deep Learning

Environmental perception is a key component of intelligent driving technology.Its key lies in making self-driving vehicles better simulate the perception ability of human drivers,so as to understand their own and surrounding driving situations and provide necessary information for decision-making and planning Stage.Autonomous vehicles driving on the road,and road environment information is essential for safe driving.Road obstacle detection is an important research content of the direction of the environment perception for intelligent driving.Since the driving vehicles generally have a high speed,the self-driving vehicles should detect the road environment in advance as much as possible,which requires the environmental perception system of intelligent driving vehicles to have the ability to detect the road obstacles at a long distance.The research in the field of object detection has been rapidly developed due to the rise of deep learning technology.Many excellent algorithms for object detection based on deep learning have emerged.These excellent algorithms have greatly improved the speed and accuracy of object detection,But the existing object detection algorithm is not satisfactory for detecting small objects.In addition,there are few researches on long distance small object detection algorithm on structured road in the field of object detection.The working mechanism of object detection algorithm based on deep learning and the current mainstream object detection algorithm is studied in this thesis.Since this study needs a large amount of data to train the CNN models,and the existing dataset related to this study are too small to meet the requirements of the experiment,this paper collects the largest dataset of this research direction in the real environment and virtual environment respectively,and makes detailed annotation.In addition,inspired by VPGNet proposed by Seokju Lee et al.,A road vanishing point detection model RVPGNet based on global features is proposed in thesis.The detection accuracy of RVPGNet model in the test dataset was 90%and 99%,respectively,when the distance from the groundtrueth of vanishing point was within 64 and 128 pixels,which met the experimental requirements.In this thesis,the road vanishing point detection network is integrated into the target detection network to obtain a new small object detection model LDRODNet.This model takes the road vanishing point as the prior knowledge to find the region of the object in the image,and then detect the region,which improves the detection accuracy and reduces the computation.The AP_S obtained by LDRODNet in the test dataset is about 4 percentage points higher than the current best object detection network.