On-road Vehicle Detection Based On Multi-feature Fusion In Video Sequence

On-road Vehicle Detection, lane detection pedestrian detection and obstacledetection together constitute component of Vehicle Active Safety systems. There arestill a lot of challenges about real-time requirement and robustness. The currentmainstream method used for on-road vehicle detection is based on statistical learning.Cascade classifier based on Adaboost algorithm exists obvious defects: detectionslowly and relatively low detection rate. These articles therefore improve the overallperformance of the traditional cascade structure. The main cont ributions of this papercan be outlined as follows:Firstly, a method combined with BRIEF and Haar-like features is proposed toenhance the discriminative abilities of the feature set and accelerate the detection.Different weak classifiers for BRIEF features and Haar-like features are designed forworking together to construct a strong classier.Secondly, BRIEF and Haar-like features are improved. Haar-like feature isimpoved by adding a new rectangle feature to use the feature of the vehicle shadow.Remove the four kinds of vehicle detection less used to the traditional Haar-likerectangle features. BRIEF features for noise-sensitive shortcomings, with a small areato replace a single pixel and match pixel matching, and the introduction of integralimage pixel and a small area to speed up the calculation; against BRIEF scaleinvariance does not have the characteristics of the shortcomings, the introduction ofscale space theory solve BRIEF feature points at different scales match.Additionally, different weak classifiers for BRIEF features and Haar-likefeatures are 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.Finally, using a vehicle tracking algorithm based on Kalman filter accelerates thespeed of vehicle detection. Scale and position of the vehicle as a vector of the Kalmanfilter to predict the next frame of the vehicle scale and location of the original and thepredicted vehicle position of the vehicle are identified as a region containing thevehicle, thereby reducing the search time of detection.