英文摘要 | Traffic Data Collection System, as a part of Intelligent Transportation System (ITS), plays a significant role in traffic management and control, transportation planning and many other applications. Traffic data collection builds upon the detection of on-road vehicles; here, video detectors offer many advantages over traditional inductive loops. Although the field of video detection has been extensively explored in the past few years, due to the complexity and diversity of traffic scenes, many existing algorithms are too simple to satisfy the speed, accuracy and robust demands in practice. This thesis aims to develop efficient methods in solving some key problems. Video-based traffic data collection relates to many common research topics; including road camera calibration, moving object detection, moving shadow removal, traffic parameter extraction, embedded system development, and so on. In this thesis, we not only study some related fundamental theories, but also develop an embedded traffic data collection system for practical applications. The main contributions of this thesis are as follows: 1. We proposed a simple but convenient approach for camera calibration in traffic scenes. With the equation array of camera parameters as the kernel, various prior knowledge corresponds to various camera parameter recovery methods, making this approach very practical. Besides, experiments prove that different camera parameter recovery methods are essentially equivalent. 2. We discussed the moving object detection approach by combining three-frame difference and background subtraction. In particular, we propose an 8×8 pane-based connected component extraction method, which is very effective in improving the connectivity of moving objects and the adaptivity of algorithms in dynamic environments. 3. We proposed a spectrum-based moving shadow removal approach. First, through automatic and on-line learning of pixel-wise shadow, we calculate the shadow color model; then, combing the shadow color model and the shadow edge model, we obtained good shadow recognition effects; finally, we perform spatial analysis in order to eliminate the penumbra and further reduce the false positives and the false negatives of the shadow. 4. We studied the virtual loop-based traffic data collection approach. Using the pixel motion and edge features in the virtual loop, we propose a novel vehicle detection method, which is not sensitive to a wide variety of artifacts. Experimental comparisons with a state-of-the-art method prove that the proposed method has low computation load, high accuracy, and a strong robustness. 5. We designed and implemented a DSP-based embedded traffic data collection system. We obtained a large amount of intersection traffic data, which initially confirmed the accuracy and robustness of the system working in traffic scenes over a long time span. |
修改评论