Real-time Road Vehicle Identification Method Based On The Haar-like Features

Real-time on-road vehicle detection is a key technology in many transportation applications,such as driver assistance,autonomous driving and active safety.Due to varieties in vehicle appearance and complexities of outdoor environment,it is important to distinguish vehicles from other objects as robust and reliable as possible.The supervising approach of statistical pattern recognition is utilized in this thesis to train a classifier based on Haar-like feature which implements real-time on-road vehicle detection.Based on previous research,the structures of Haar-like rectangular features and integral image that can be used to compute the value of features rapidly are analyzed. Eleven sorts of expanded rectangular features that Lienhart R.put forward are adopted to represent vehicles.In addition,this paper develop a single feature to represent the shadow at the bottom of vehicles,and construct U type feature using two vertical edges and shadow of vehicle.Haar-like features are used to construct weak classifiers,and Gentle AdaBoost algorithm is chosen to train the strong classifiers.Then,the real-time on-road vehicle classifier based on tree structure is constructed by combining the strong classifiers,which also uses clustering-and-splitting step recursively,and the root node of tree classifier is a strong classifier that employs U type feature.Multiple detection window approach is introduced to detect vehicles in images, avoiding scaling image directly in traditional way.Tree classifier rapidly rejects as many negatives as possible at the earliest layers,only true positive areas reach the leaves terminal strong classifiers.The data sets which consist of training images and testing images are collected,and video data set is collected for real-time vehicle detection.The real-time on-road vehicle classifier based on tree structure is trained using training data set and evaluated in testing data set and video data set.Experimental results show that the tree structure classifier is better than cascade classifier in both detection accuracy and computational efficiency.