Developing Visual Intersection Vehicle Detection Model at Night Applied on Neural Network

• Main Authors: Wen-Fu Lai, 賴文復
• Other Authors: Chun-Hai Fan
• Format: Others
• Language: zh-TW
• Published: 2004
• Online Access: http://ndltd.ncl.edu.tw/handle/64029579186909101526

碩士 === 淡江大學 === 運輸管理學系 === 92 === Because of video image detector can obtain more traffic parameters than traditional ones, and provided with many advantages, including cheap cost, easy installation and quick maintenance etc. So it is suitable and efficient to extract traffic parameters of image processing. Due to insufficient illumination and instable illuminant at night, previous researches almost focus on daytime and cloudless weather. It is few researches study on the environment of night or intersection. Therefore, our research will study the circumstances of intersection at night. The feature segmentation is a critical technique to decide that night detection success or failure. For example, when vehicle was drove at night, the most obvious features included lamp, the light reflection of car and the light reflection of ground etc. However, owing to the vehicle behavior in intersection are more complicated than in road. Hence, how to segment features accurately from complicated intersection environment at night, which is the first arduous problem we must to overcome. In this problem, we developed a neural segmentation method based on back-propagation. Moreover obtain vehicle features by neural segmentation method, which result manifest better performance than heuristics rules and statistics method. Besides, the instability of illuminant at night and the turning behavior of car in intersection are caused the variation of vehicle’s features. So, the second difficult problem is how to cluster the features with the same vehicle. In order to solve this one, we adopted fuzzy system to build the fuzzy block clustering method integrating the most part of features. Performing these clustered works by similarity between blocks, and the result manifest okay performance when deal with low- density traffic flow. In our study, we successfully extract the traffic parameters are included vehicle classification and turning traffic flow. After our experimental analysis, the accurate rate of identifying big vehicle is 58.6 %, small vehicle is 85.5 %, motorcycle is 84.5 %；The accurate rate of detecting transferred traffic flow, in south-north phase is 82.5 %, and in east- west is 83.9 %