Research On Vision Based Front Vehicle Detection Algorithm

Along with the rapid development of social economy, the car ownership is rapidlyincreasing, but the resulting road traffic accidents are also rising year by year. Vehicledriving safety gets more and more attention of people. In order to avoid the occurrence oftraffic accidents, reduce the personal injury and lower the degree of the loss caused bytraffic accidents, the global auto manufacturers, suppliers and scientific research institutesare giving more and more attention to related research in the field of vehicle assistancedriving.Automotive vehicle detection system is quite an important part of vehicle assistancedriving system. Goals of on-board vehicle detection system are mainly detecting othervehicles around itself by installing the on-board camera in their own vehicles and whenthere is a possible collision with other vehicles, it will remind drivers to avoid trafficaccidents as quickly as possible.On road vehicle detection algorithm is studied in the thesis. The vehicle detectionalgorithm follows two steps: vehicle hypothesis regions generation and vehiclehypothesis verification.In the vehicle hypothesis regions generation step, shadow underneath a vehicle isused as a sign pattern. First, a dynamic threshold for segmenting the shadow in the imageis decided by approximating the mean value and the standard deviation of the road grayvalues in front of self-vehicle. In the shadow extraction step, two methods which arebased on the shadow’s rectangular-like feature and edge features of the bottom of shadoware combined together. And finally vehicle hypothesis regions are generated according tothe location of shadow.In vehicle hypothesis verification step, preliminary screenings of the vehiclehypothesis regions are conducted by the vehicle’s edge symmetry characteristics firstly.And then HOG (Histograms of Oriented Gradient) and linear SVM (support vector machine) are used for vehicle hypothesis verification. After computing HOG feature,linear SVM is used to train the classifier. Finally, the trained classifier is used for vehicleclassification.