Vehicle Detection Algorithm And Scene Adaptive Analisis Based On Co-training Learning

With the sustained and rapid development of Chinese society and economic, aswell as the rapid increase of the number of motor vehicles, urban transport problemshave become increasingly severe. In order to improve the level of traffic managementand the operating efficiency of traffic, researchers have proposed the concept ofintelligent transportation systems. In the intelligent transportation research, the vehicledetection as the basis of the traffic management has very important theoreticalsignificance and application value.Due to the diversity of the external appearance of the vehicle and the complex ofthe traffic background, it is difficult for video-based detection of vehicles toaccurately achieve all-weather vehicles. Machine learning is one of the effectivemethods for vehicle detection. But there are still two questions: how to obtain asufficient number of training sample images, and how to adapt the traffic scene. Thethesis conducts the research of the two issues in depth. The main work andcontribution of the thesis include the following aspects.1. The vehicle detection method based on co-training. For solving the difficulty ofobtaining a sufficient number of training sample images when detecting the vehicle,by using machine learning method, the thesis propoes a framework of the co-traininglearning algorithm. According to the requirements of the algorithm to the featurespace, the thesis uses AdaBoost classifier based on Haar feature and SVM classifierbased HOG feature as the mutual supervision classifiers, to predict the labels ofunlabeled samples and then add them to the other classifier’s sample library.2. The co-training learning-based scene adaptive algorithm. When using aparticular classifier trained for a traffic scene to detect the vehicles in the other trafficscene, the detection accuracy of the vehicle will greatly reduce. For this problem, thethesis uses co-training learning algorithm to adapt traffic scene in vehicle detection.On this basis, the thesis proposes a similarity judgment method of traffic scenes todetermine the appropriate use of the method for the scene adaptability. At the sametime, the thesis proposes a label confidence analysis method to improve the predictionaccuracy of the labels in the semi-supervised learning process.3. Development of a system platform for verifying the research of vehicledetection and scene adaptive algorithm. The system platform implements training,testing and detection process of AdaBoost classifier based on Haar features and SVMclassifier based on HOG feature, including the four parts of establishing samplelibraries, training classifiers, testing classifiers and the target detection.4Verification of the vehicle detection and scene adaptive algorithm basedco-training learning. By using the system platform developed, the thesis establishs theinitial training sample database, trains the classifiers, tests the classifier performancebefore co-training learning and after learning, and verifies co-training learning methodin the situations when increasing the number of training sample images as well as thetraffic scene adaptive algorithm.