Research Of Video-based Vehicle Detection Algorithm For Driver Assistance System

Although the rapid development of automobile industry has brought our life much convenience, the increasing traffic accidents caused by rapid growth of automobile production make driving safety problem more serious. Therefore, the research of Intelligent Driver-Assistance and Safety Warning system (IDASWs) has become an important subject of the research of intelligent traffic system both at home and aboard.The focus of the research into Intelligent Driver-Assistance and Safety Warning system (IDASWs) is how to efficiently detect the vehicles on the roads in order to alert drivers to the driving environment and possible collision with other vehicles. Vehicle images always contain lots of horizontal and vertical structures. Starting with the captured structure of vehicle images, this paper first extracts the region of interest and then by the verification of which gets target vehicles. The content of this paper can be divided into the following two parts:1,The hypothesis of candidate vehicle location. This paper proposes a multi-scale edge analysis approach based on distance information. Firstly, according to the distance information, three sub-images of three-scale are obtained, whose horizontal structures are then analyzed. By extracting baselines of vehicle location of each sub-image and combining the vertical edges analysis, initial vehicle location is attained. Considering the various degrees of collision risk caused by the vehicles in different distances, this paper attaches a distance risk factor to each candidate vehicle location for the final verification.2,The verification of candidate vehicle location. There are some fake targets in the candidate vehicle location, so it is necessary to identify real vehicles from those hypothesis. Verifying a hypothesis is essentially a classification problem of two patterns, i.e. vehicle and non-vehicle. By applying the method of machine learning, certain amount of vehicles and background sample are extracted. Then this paper proposes an improved Gabor Wavelet feature extraction method and applies Posterior Probability Support Vector Machine (PPSVM) to train the classifier. Firstly, extract features from the samples by our method and put the feature vectors into PPSVM to learn the decision boundary. During the verification of candidate vehicle location, feature vectors from those locations are extracted and put into the classifier which has been trained to get the probabilities that those locations are vehicles. And finally combined with their distance risk factor, false targets will be removed and target vehicles are got.The system has been tested in some sequences of videos under different traffic conditions and it has exemplified good performance in real time.