The Research Of Moving Vehicles Detection Technology Based On Machine Vision

Video based moving vehicle detection is the fundamental part of Intelligent Transport System. Its performance has a direct influence on semantic information understanding such as vehicle tracking, vehicle recognition and behavior analysis. Moving vehicle detection can realize traffic parameter extraction and further automatically detect traffic incidents. However, in the out-door vehicle detection system, detection result can often be affected by some outside factors such as camera disturbance, light variation and leaf shaking. How to detect and extract moving vehicles in a real-time accurate way in outside complex environment is the research topic of this thesis.Aimed at the limitation that background subtraction method is liable to be affected by complex outside environment, this thesis provided a modified K-mean clustering algorithm to model the background in complex environment. Multi-mode influences such as noise and camera disturbance in the outside environment are described by K sub-clusters. Observed values of every pixel are clustered by comparing the distance between each sample and the sub-cluster center at that pixel. The cluster numbers are determined during the clustering process. After training for some time, background model is composed by sub-cluster with the maximum samples. This algorithm can beautifully extract the background even in the video where vehicles are always moving.For the situation that vehicle in the background suddenly starts to run or vehicle suddenly runs into the current background, this thesis firstly detects the variation area with background difference and then detects the edge contour of the moving vehicle in current image using canny operator. Finally, the above two algorithms are associated to advance the robustness of the system, restrain the appearance of the false targets and promote the accuracy of the detection.Considering that sun light in outside environment will cause shadow on the vehicle, this thesis detects and reduces the shadow using HSV space, which gets a good performance. This method further improves the accuracy of moving vehicle detection.Aimed at the problem of traditional snake algorithm relaying on the original contour, this thesis uses the foreground vehicle boundary from modified algorithm rough segmentation as the original edge contour of the modified snake algorithm, which reduces the human intervention when snake model is used to segment complex image. Vehicle segmentation method based on active contour Perona-Malik algorithm is used to overcome the limitation that traditional method cannot enter the dent of the vehicle image. Besides, this algorithm is seldom depend on the original contour and its robustness is satisfactory.Finally, experiments are conducted using live traffic video images, which gets a good performance and tests the effectiveness of the algorithms.