Research On Vehicle Feature Extraction And Recognition In Complex Condition

With the development of society and the increase of vehicle numbers,Intelligent Transportation System(ITS) plays a more important role in transportation management and gets more attention. Vehicle feature extraction and recognition is very important in ITS. This paper proposes improved algorithms based on feature extraction and feature matching and applies them to vehicle detection and tracking. The main works of this paper are presented as following:(1)SURF(Speed-Up Robust Features)is a robust and fast descriptor for many applications. However neither can it detect symmetrical features, nor can it consider global context. This paper proposes a new algorithm that combines symmetrical SURF with global context. It enables SURF to detect symmetrical matches through mirroring transformation and reduces mismatches when local descriptors are similar. The proposed algorithm is used in vehicle detection. Experimental results show that symmetrical SURF with global context improves the accuracy of feature matching.(2)Because the commonly used feature matching methods ignore geometric information in images, the efficiency and accuracy of matching are not perfect. Although BP-SIFT(Belief Propagation) method uses geometric information, the distance among key points isn’t always the same. Neighbor points selected based on the distance between feature points can only represent partial geometric information of the image. And the algorithm costs too much time. The ratio of the distance between the key point and a neighbor point to the mean distance between the key point and all neighbor points is proposed as new constraint. BOW(Bag-of-Words) model is used in selecting neighbor key points. Experimental results show that the proposed method improves the accuracy of feature matching and reduces the time complexity.(3) The improved symmetrical SURF and feature matching methods are applied to vehicle detection and tracking.Searching for areas with high symmetry to locate vehicles from improved symmetrical SURF matching points is uesd in vehicle detection. The proposed feature matching algorithm is used to searching for matching points between images in vehicle tracking.Experimental results show that both algorithms improve the accuracy of vehicle detection and tracking.