Traffic Scene Understanding Based On Image

Abstract:Currently, the video image processing of traffic scene has become a hotspot of computer vision. Focusing on the detection of traffic areas, the superpixel segmentation algorithm of traffic scene images, the superpixel area description of traffic images and the semantic understanding of traffic scene images, the research of this paper are reflected in the following aspects:Because there is no fixed pattern for road area shape, no consistent color, as well as influences of noisy, light and shadow, this paper puts forward a detection method of traffic areas based on improved Adaboost. The method enhances performance of the classifier to identify roads.For the large data amount of traffic videos and images, as well as the big calculated amount, this paper introduces the concept of superpixel and comes up with advanced SLIC superpixel segmentation algorithm to process traffic image segmentation. Compared to traditional over-segmentation methods, the improved SLIC algorithm solves some problems such as uneven shape of superpixel and poor boundary information to certain degree.A covariance traffic image pixel formulation based on Riemannian manifold is proposed to integrate the characteristics of pixel’s color, gradient and texture effectively, reduce the dimension of features, and increase the accuracy of the traffic image description. Besides, bag-of-words representation is presented, which can map the low-level image visual features to high-level semantic features. The representation maps the local low-level image features of traffic image to visual word, and it is showed by a histogram that is available with classification features.Since traditional unsupervised LDA topic model’s classification effect is poor, this paper presents semantic classification and labeling method of traffic road areas, which is based on SLDA probabilistic model. This method makes use of prior knowledge of traffic scene, integrates the context relations of entire traffic scene topic and local object effectively, and improves the classification performance of traffic scene image to get a more accurate estimate.Experimental results shows that methods proposed in this paper can improve semantic analysis of traffic scene image.