Research On Detection Method Of Motorcycle With Additional Components On Unstructured Road

Autonomous driving of intelligent vehicle is an important way to reduce traffic accidents and ease traffic congestion.Autonomous driving systems are generally divided into four parts: environmental perception,navigation,motion control and user interaction.Accurate perception of road traffic environment is the basis and prerequisite for automatic driving system.Realizing accurate perception of road traffic environment is of fundamental importance for autonomous driving systems.It is common for moving targets to rush into the road and cause traffic accidents due to the lack of clear road signs and boundaries on unstructured roads such as urban non trunk roads and rural streets.In addition,the number of motorcycles and electric motorcycles is greater than that of cars on unstructured roads.According to statistics,the proportion of traffic accidents caused by motorcycles is the highest every year.Nowadays,some motorcycles are often equipped with windproof quilts in autumn and winter,and most takeout employees install hanging boxes on their motorcycles,which bring challenges to the existing visual model of target detection.Taking the environmental perception of intelligent vehicles as the research object,this Ph.D.dissertation makes an in-depth study on the problems existing in the existing road recognition and target detection algorithms,and proposes a visionbased motorcycle with additional components detection model(VACDM).It can effectively solve the problems of insufficient samples of motorcycle common hanging box,reducing accuracy of motorcycle detection by adding components,and predicting and tracking targets in a certain area beside the road.The model includes road region classification model(RCM),dual backbone instance segmentation network(DISNET),and image editing algorithm for hanging box.The main contents of this paper include a few aspects:(1)An image editing algorithm based on 3D model was proposed.Most of the collected hanging box samples are takeout boxes,and the sample of ordinary hanging box is few,which is not enough to train a segmentation network with good performance,we propose an image editing algorithm based on 3D model.The 3D motorcycle model is fitted to the motorcycle in a real image,and the aligned 3D model is used to edit the real motorcycle image directly to generate effective training data.The algorithm can solve the problem of lack of training samples in the ordinary hanging box.In order to test the quality of the samples,a comparative experiment was carried out on the effects of the samples generated by different methods on the network training effect.The experiment show that accuracy of the network segmentation can be improved by using the combination data set training.90% of the samples in the combined data set were generated by image editing algorithm,and the rest 10% of the samples were randomly selected from the real data set.(2)A dual backbone instance segmentation network(DISNET)was proposed.Traditional multi-target detection algorithm generally only considers convolution feature based on the overall semantics,and integrates additional component features,we propose a dual backbone instance segmentation network.The network divides output results of motorcycles into three categories: installation of windproof quilt,installation of hanging box and no installation.The self collected data set is used for training and testing,and the comparative test between DISNET and other high-level algorithms was carried out.The experimental results show that DISNET can realize real-time multi-objective instance segmentation on a single graphics card,and its performance achieves state-oftheart performance in data set testing.By introducing deformable convolutional networks(DCNs),we improve backbone network in DISNET to improve the accuracy.On the premise of ensuring the real-time detection of VACDM,we conduct exhaustive experiments on the replacement scheme of DCNs,and find the most suitable replacement scheme for DISNET.(3)A new road classification model(RCM)was proposed.Targets in a certain area beside the road have an important impact on driving safety,this impact is more obvious on unstructured roads and is more prone to traffic accidents,we propose a new road classification model(RCM).RCM divides an image into four parts: driving region,roadside region,cross region and other region,and implements different safety strategies for each region.On this basis,the field experiments on the actual road show that the algorithm can track and predict the targets in the driving region and roadside region,and provide timely warning information to the driver before the target rushes into the road.