Research On Vehicle Detection Methods Based On Stereo Vision And Machine Learning

In recent years,the traffic accidents have become a wide spread concern in society.The loss of personnel and property caused by traffic accidents has seriously affected our life.In order to improve safety and reliability,the major automobile manufacturers have started to do research on active safety driving technology and intelligent vehicles.The concept of intelligent traffic system has also been proposed accordingly.The front vehicle detection has always been an important problem in the research of intelligent traffic system.So far,research on vehicle detection has relied on various sensors.Compared with laser radar and millimeter wave radar,the vision sensors have the advantages of low cost and abundant information.In recent years,with the improved hardware performance,the vision sensors are used more and more widely.Vision-based vehicle detection methods can be mainly divided into monocular vision method and stereo-based vision method.Compared with moncular vision,stereo-based vision method can not only provide abundant appearance information,but also provide the depth information,which can achieve a higher precision.In our paper,a hybrid method based on stereo vision for real-time vehicle detection in urban environment is proposed and it can be divided into two steps,hypothesis generation and hypothesis vertification.There are two steps in the hypothesis generation,generating the initial candidate regions and confirming the candidate regions.Firstly,we propose a shadow-based vehicle detection method.The candidate areas where vehicles may exist are generated by extracting the underearth shadow of the vehicle.Then,to determine the candidate regions furtherly,we calculate the disparity map according to the stereo matching algorithm,obtain the depth information and the actual size of vehicles.Finally,we cluster the disparity map to eliminate noises,always with obvious depth variation in the disparity map.In the hypothesis verification,we make an existence verification to the candidate regions.In this paper,we combine the methods of the histogram of oriented gradient(HOG)and the support vector machine(SVM)to identify the targets.In order to further improve the accuracy of the experiment,we lead in a multi-scale classifier to optimize the experimental results.In the experiments,we test on two datasets.Our real-time vehicle detection method can achieve a 95.9% accuracy rate on outself data in the urban environment.We also select 8 video sets from the open source dataset KITTI to test our method.The accuracy rate is over 90% and the average accuracy rate is 92.95%.In order to verify the accuracy of the detection distance,we also make some tests.The deviation is positively correlated with the detection distance,the minimum deviation is 2.7%,and the average deviation is 4.9%.The experimental result shows that the vehicle detection method based-on stereo vision and machine learning can effectively identify the target vehicles in various urban scenarios in real-time.At the same time,the clustering algorithm and machine learning algorithm can remove noises,such as trees,guardrails,signages and buildings effectively.With the study of the intelligent traffic system in deep,our method can also be applied to Advanced Driver Assistance Systems(ADAS)on the collision warning system and active braking system for front obstacle detection,which has a good application prospect.