Research Of On-road Vehicle Detection Based On Multi-feature Cascade Classifier

On-road Vehicle Detection is a main component of Vehicle Active Safetysystems. Vision-based Road Vehicle Detection is a hot spot in the research of On-roadVehicle Detection systems. But there are still a lot of challenges about real-timerequirement and robustness that the vision-based vehicle detection systems mustencounter. The current mainstream method used for on-road vehicle detection is basedon statistical learning. Cascade classifier based on Haar-like feature and Adaboostalgorithm exists obvious defects: training time is too long, detection rate is lower thansome other classifiers, and it is easy to cause over-fitting while using large number offeatures in the rear layers. This article therefore improve the overall performance ofthe traditional cascade structure.The main contributions of this paper can be outlinedas follows:Firstly, in order to enhance the discriminative abilities of the feature set, amethod that combine HOG features and Haar-like features is proposed. The number ofHaar-like features is reduced, then combined with modified HOG features. This twokind of features are calculated quickly using integral image and integral histogram.Additionally, different weak classifiers for HOG features and Haar-like featuresare designed. Aiming at reduce the the number of features used in the cascadestructure, real-valued weak classifiers are utilized instead of binary weak classifiers,then Gentle Adaboost algorithm is adopted to train the layer classifiers.Fainally, based on the fusion features, a cascade classifier combined withSupport Vector Machine is proposed. In the rear layers of the cascade, feature vectorscomposed by the features that selected by Gentle Adaboost algorithm are used to trainrobust SVM classifiers, saving the training time of the cascade structure and avoidingover-fitting.Taking advantage of OpenCV and C programming language under theprogramming environment VS2008, the proposed cascade classifier for on-roadvehicle detection is trained. Compared with traditional cascade classifier,experimental results indicate that the cascade classifier proposed here has higherdetection rate while saving the training time.