Research And Implementation Of A Vehicle Detection Algorithm Based On Monocular Vision

Vehicles bring people’s lives a lot of conveniece, meanwhile a lot of serious problems too. In the safty aspects, traffic accidents causing a large number of casualties and economic losses each year. To solve these problems, many countries put a lot of manpower and material resources to the development of Intelligent Transportation Systems (ITS). Driver Assistance System (DAS) is an important research direction in ITS.The monocular vision based vehicle detection algorithm this thesis reserched is used in. DAS, according to the traffic scene images from a monocular camera, detected if there are vehicles or not in the front area rapidly. Then the test results are sent to the driver and other modules of DAS. It can help the driver to collect traffic information and remind the driver potential threats in the front, provides the possibility of vehicle collision warning and intelligent driving.This vehicle detection algorithm is a moving vehicle detection technology. Since DAS require a higher real-time performence, while the vehicle hardware can not be very high-powered, the algorithm design of moving detection is more difficult than the static detection. This paper analyzed many existing vehicle detection algorithm, proposed an improved scheme against the shortcomings. To avoid searching the whole image to find the vehicles, the algorithm is divided into two stages:the shadow feature extraction stage and the vehicle feature recognition stage. In the shadow feature extraction stage, use the features of shadow underneath the vehicle to segment image and generate ROI; in the vehicle feature recognition stage, use Haar-Like Feature to construct weak classifier, and use the Adaboost algorithm to improve the weak classifier to strong classifier. Then use the strong classifier to recognize ROI and obtain the results.Experiments show that the algorithm can detect the vehicle on-road in the front with a higher recognition rate, and the probability of misrecognition decreased; real-time performence is also in acceptable range.