Research On Vehicle Detection Algorithm Based On Image Recognition

With the rapid growth of today's social economy, social demand for transport has increased and traffic density also became bigger than before. In the daily life, traffic accidents occur frequently and they cause great damage of personal and property safety. In recent years, in order to improve the situation of traffic safety, prevent the accidents,countries all over the world increased investment for traffic system management, and gradually formed a special research field of road traffic management. According to figures from the traffic accident, the major threats that driver facts is from the other vehicles. As a result, a vehicle system that could help the driver master driving environment has attracted widespread attention. Now the vehicle system camera installed on the vehicle rather than in highway monitoring system. In these systems, vehicle identification with robustness and reliability is the first step.But the vehicle has a great variability in the appearance of classification, in the shape,size, and color are different. The appearance of the vehicle dependents on its location, and it is easily affected by complex external environment. Therefore the vehicle detection using an optical sensor is still a big challenge. With the continuous innovation of the current image processing and pattern recognition technology, computer vision is more and more used in vehicle recognition. As the vehicle speed is closely related to the processing speed, vehicle system requires faster processing speed than other applications. But vehicle detection is same as most object detection, there exist the problem of large amount of calculation and slow detection, so that it cannot become a utility program. How to improve the detection algorithm, and realize the real-time detection is still a difficult problem in current study.The main work of this paper is to study the three features commonly used at home and abroad, namely Haar, HOG, LBP, and Adaboost learning algorithm. Based on these three characteristics, we train the Adaboost classifier and compare the training time, target identification result in the vehicle detection algorithm. Finally we obtain the vehicle recognition classifier of high performance, thus improve the practicability of the vehicle detection algorithm.In the stage of training classifier, the three features are different. From the sensitive degree of the classifier on sample quality, the most sensitive characteristic is Haar classifier.Therefore, Haar characteristics can be used to determine the sample quality according to the training process, to avoid unnecessary training process. From the whole training speed of the classifier, LBP feature classifier is the fastest, and in same sample, the training series of LBP feature classifier is the least. From the time of classifier training to extract feature, the pre-calculation time of Haar features is the longest. In terms of quality of the classifier,HOG classifier is optimal. It has good detection rate and lower false detection rate on the vehicle. From the target recognition time, recognition algorithm based on HOG feature is the fastest.Because the accuracy of vehicle detection algorithm affect the traffic safety, therefore,it is reasonable to sacrifice classifier training time to get higher accuracy. What's more,considering of the practicability of the vehicle detection algorithm, we need ensure that the real-time performance must be very good. Therefore, taken together, the HOG feature is undoubtedly the best choice for vehicle detection algorithm in the target feature extraction.