Study On Image/Video-Based Vehicle Detection

Intelligent Transportation System (ITS) is an important application of intelligent video monitoring system in the field of modern traffic, in which video-based car detection represents the future trend for the traffic management. Generally image/video-based car detection can be viewed as a category of classification problem composed of sample set contruction, feature detection, and classification model training, etc.Based on the state of the art for car detection, this thesis focused on the research work from such aspects as the construction of traininig set, feature detection, and classification model training. The main contributions can be summarized as the following.(1) Training set construction based on multi-scale samples.In the phase of training set construction, multi-scale positive set is produced by sampling and interpolation, while the negetive sammples are gathered by bootstriping strategy.Experimental results show that the classification performance based on multi-scale sample set training outperforms the traditional method.(2) Sample description based on MbSBP feature detection.By introuducing a kind of new feature descriptor called MbSBP (Median filter based Star Binary Pattern), the texture and structure information can be made full use for car object detection. Experiments show that the performance of object detection can be improved by means of MbSBP feature.(3) Classifier based on SC-Adaboost model.Although AdaBoost-based object detection from image/video dada holds the characteristics of good detection precision with high detection speed, the training procedure is much more slowly especially when the number of both samples and feature dimensionality is high. With the aim of efficiently improving the traing performance, this thesis proposes an algorithm called SC-AdaBoost. Experimental results for vehicle detection and pedestrain detection show that when the number of training samples is very large, the proposed algorithm can evidently reduce the whole training time without loss of detection performance. (4) Video car detection based static cameras.By combining background substraction and the above research work, a video-based cardetection framework is finally realized in this thesis. Experimental results show that the proposed framework performs well.