Multi-View Vehicle Detection Technology Based On Adaptive Boosting Algorithm

Vehicle detection is an important research theme in the topic of Pattern Recognition, Image processing and Computer Vision; it has a wide application prospects and practical value in many fields such as video surveillance, content image and video retrieval, automatic vehicle recognition and artificial intelligence, etc.Given an arbitrary image, adopt certain algorithm or strategy to search in order to determine the existence of vehicles. If exist, then return the position, size, view of vehicle. As in real life with different perspectives vehicles often appear in the video image, and in order to improve the robustness of detection, we have to consider different appearance which the vehicle in a variety of complex backgrounds, different direction, angle, scale and other circumstances revealed. Namely, Multi-View Detection. In this thesis, the major work as follows:(1) In the aspect of feature extraction and detection, in order to improve the compute speed of feature, use Harr-like feature to presentation image and introduce concept of "integral image". Besides, to improve face detection rate, use AdaBoost technique to choose features for compose strong classifier and applying "Cascade" strategy for detection. AdaBoost algorithm is a technology which viola applied in face detection. This method, obtain good detect performance and realize the face of real-time detection. It basically achieve real-time. Therefore, it can be used in Multi-View Vehicle detection and practical application.(2) In the aspect of construct various perspectives of vehicle detector, we adopt Haar-Like Features and AdaBoost learning algorithm to construct the various perspectives of the vehicle detector. Use cascade method to constitute a strong classifiers various perspectives of the final classifier in the training multi-view vehicle detector. Finally, in the testing phase, introduce view estimate to predict five different perspectives, then comprehensive test results and obtained experimental results.(3) In the aspect of increase training samples, to solve the problem of insufficient training samples, we introduce a mechanism to increase the training samples. Create training samples from one image applying distortions, then generate a few training samples, and then iterate this process. Therefore, we can get thousands of training samples and solve the problem of insufficient training samples. Experimental results show that increasing the training samples can improve detection.This thesis adopts Haar-Like Features to presentation image and use AdaBoost algorithm to construct strong classifiers. Besides, use cascade method to constitute classifiers. Finally, in the testing phase, introduce view estimate to detect and obtain experimental results. The results shows that adopt Haar-Like Features and AdaBoost learning algorithm can solve the problem of vehicle multi-view and achieve the destination of detection.