Research And Implementation Of Low-Level Semantic Extraction In Intelligent Transportation Scene

In recent years, intelligent transportation system (ITS) has been widely used and a large amount of surveillance cameras along roads record transportation scene in real time. How to use these surveillance video data and extract semantic of transportation scene is a key to build ITS. Vehicle logo recognition, vehicle color recognition and vehicle speed deteticon are the key issues in semantic extraction of intelligent transportation scene. To solve these problems, we design and implement an intelligent transportation scene low-level semantic extraction system.In the aspect of vehicle logo recognition, we fisrt utilize the relative position relationship between licence plate and logo to coarsely localize vehicle logo and then use morphological method to accurately determine vehicle logo position. We propose a logo recognition method based on multiple features fusion, PCA dimension reduction and BP neural network. In the aspect of vehicle color identification, we first use CrabCut algorithm to segment vehicle foreground and then enhance image contrast based on luminance information. We also design a set of color decision method based on SOM neural network. In the aspect of vehicle speed detection, we first utilize GMM to detect the vehicles which drive into the detection area and track those vehicles to obtain the vehicle location and size information in the surveillance video. Due to the lack of both internal and external parameters of surveillance cameras, we propose a novel integral-based method to map video frame distance into real space. This method can calculate vehicle speed without camera parameters. The experiment shows that our system has high recognition accuracy and good real time performance...