Vehicle Detection And Recognition Based On Harris-SIFT And Normalized Cuts Algorithm

Intelligent Traffic System (ITS) is an important part of urban security system.Traditionally we use radar or ground loop to get real-time traffic information and vehicleevents. On the one hand, due to the reason that traditional means can provide less information,it will result in a lot of mistakes; on the other hand, the necessary transformation on the roadwill introduce a huge infrastructure costs. This paper mainly focus on vehicle detection incomplex outdoor environment, especially on how to extract Harris-SIFT operator from theinput video image, Segment the track graph with normalized cuts and realize system on DSP.The ITS systems designed is a precautionary integrated system, which can provide decisionmakers with an intuitive, accurate, real-time and rich traffic information. It has been widelyused bayonet, highway traffic monitoring applications.Feature points extracted from the traffic monitoring video images can easily beinfluenced by environment illumination changes, scale changes and rotation during vehiclemove. In order to extract stable image feature points, we extract the image feature points withHarris algorithms, establish128-dimensional scale-invariant feature vector in the feature pointneighborhood which is borrowed from SIFT algorithm ideas, calculate the main andsecondary direction of the neighborhood and normalize eigenvectors. The algorithm extractthe feature points with Harris algorithm instead of SIFT algorithm, which reduce theextraction time and hardware consumption. Besides, it can enhance the robustness on rotationand brightness changes with the advantages of SIFT, improve efficiency and ensure real-timeprocessing and extract feature points with strong stability.In order to detect vehicle, we conduct feature points match in continuous image sequenceto establish feature points track. Then establish vehicle track graph with the feature pointstrack as a vertex, the relationship between each track as edges and partition the graph with thenormalized cut segmentation algorithm.The normalized cut segmentation algorithm focus on the overall impression of an image,and thus can be more consistent with human perceptual segmentation. Through thedevelopment of a certain criterion, this segmentation algorithm can efficiently segmentvehicles from a large number of feature points and reduce interference.Finally, compile the Harris-SIFT algorithm proposed into a library file for TI DMS6467 series DSP calls. More attention is paid to how to design IP Camera, Calculate, PC, Uploadthreads, focus on dual-core chip. The four threads constitute the main framework of the ITSsystem and coordination DSP nuclear, EMAC, EDMA modules to complete the vehicledetection.