Vehicle Type Recognition Based On Video Sequences

Automatic vehicle classification, the object of this paper, plays an important role in road design, traffic surveillance and highway charge. Because of inadequate information obtained, difficulty in mounting and low dependability, traditional vehicle detection and classification with induction loop or piezoelectricity sensor can not be widely used. On the contrary, vehicle classification teconology based on image processing, which settled the above-mentioned deficiency, can be widely used in auto charge, park lot management and highway surveillance and becomes the direction of automatic vehicle classification development. A new system based on video sequences and SIFT(Scale Invariant Feature Transform) local feature descriptor had been developed in this paper.The system advanced in this thesis can be summarized in three parts, which includes video segmentation part, feature extraction part and SVM recognization part.In the video segmentation part, I put forward a double inter-frame difference 'and' algorithm to segment the object. This algorithm can automatically separate the motion regions from background and avoid the influence of interference motion objects. Then pretreatments including vehicle inclination correcting, image smoothing and histogram equalization are used to reduce the inaccuracy brought out by the noise and the light which is too bright or too dark.In the feature extraction part, I used SIFT descriptor to represent the features of the image. This descriptor is invariant to image scale and rotation, and also invariant to changes in viewpoint and illumination. Then I proposed the method of sub-region division. This algorithm first divides the image into several sub-regions, then counts out all the SIFT points localized in every region, calculates the average value of these SIFT vectors, finally joints these vector to obtain the character vector of the image. The dimension of the character vector obtained in this way is independent of the scales of images and resolutions of the camera. So this method enjoys excellent expansibility.In the SVM recognization part, I first conducted research on several SVM multi-classification method, then structured the SVM multi-classifier for the system. The experimental results proved that this system can recognizes the vehicle type from video sequences with high accuracy in real-time.