Research On Vehicle Detection And Identification For Video At Traffic Checkpoints

With the growth of traffic and the need of public security in modern cities, more and more video road bayonet systems appear. Video bayonet systems show great potential in a variety of applications, such as in traffic supervising and criminal investigation. The functions may include detecting and identifying vehicles to help checking violate traffic regulations, hunting hit-and-run driving, tracking suspected vehicle, finding false vehicle tag. Unfortunately, these works mostly rely on human labor at present, which brings up work-load and lowers efficiency. Developed on current human face recognition methods, we have carried out research into the methods of vehicle detection and identification. Major aspects include:1. Vehicle detection based on AdaBoost algorithm. The basic idea is to extract Haar features from vehicle samples, combining with AdaBoost algorithm, train weak classifiers to strong classifiers. Concatenation of strong classifiers is applied at last to accelerate detection and increase accuracy.2. Feature characterization, extraction, and identification of the vehicle types. As images of road bayonet systems are subject to environment, and vehicles position can vary, Gabor wavelet transform and LBP operator is used to extract multi-scale and multi-orientation vehicle layout features in detail. The characteristic vector consisted of a sequence of block histogram is used to represent the feature characteristics of a vehicle type. Principal components analysis is used to reduce dimensions. The characteristic vectors with reduced dimensions then go through Euclidean distance comparison to finalize vehicle types. In addition, the recognition of vehicle dadging is applied to narrow sample set, improving the accuracy rating of vehicle type identification.3. Study on vehicle color identification by classifier cascade or combining classifiers, etc. Based on human eye sensation and large statistical vehicle sample pool, suitable color space and classification algorithms are adopted. For distinguished vehicle color types, Support Vector Machine (SVM) is used for classification; for less-distinguishable colors, priori probability is adopted, histogram is computed to get class conditional probability, and Maximum posterior probability is computed based on Bayes’ Classifier to identify color types.(SVM) and Bayes’ Classifier is then combined to form a decision tree, which is then simplified to bipartition problem. Experiment results show good identification of color types.