A Cascade Classifier Using the Adaboosting Algorithm for Real-time On-road Motorcycle Detection and Tracking

• Main Authors: Yao-teng Yeh, 葉曜謄
• Other Authors: Chin-shyurng Fahn
• Format: Others
• Language: en_US
• Published: 2014
• Online Access: http://ndltd.ncl.edu.tw/handle/6bh4gt

碩士 === 國立臺灣科技大學 === 資訊工程系 === 102 === With the rapid advancement of information technology, today computing power of computer is increased dramatically; making digital image processing technology has more extensive application in real life. For solving the traffic problems also produced the Intelligent Transportation Systems (ITS), ITS is an integrated hardware and software facilities of all kinds of transportation systems in order to achieve the automation and improve the quality of transport services, and one of the ring is an Advanced Vehicle Control and Safety Services (AVCSS), AVCSS is a combination of the sensors, computers, communications, electrical and control technology for vehicles and road infrastructure, to help drivers improve driving safety, using sensors to help drivers with visual sensory deficiencies reduce the risk which caused by improper driving behavior and negligence judgment. At the field of the sensors of motorcycle detection uses different technology to achieve, such as acoustic, laser, radar and camera, etc. Using laser radar can obtain a high detection rate, and can get information about the target, such as the target distance, azimuth, height, and speed, etc. But its expensive hardware also is a burden. The detection methods using the camera, because the inexpensive hardware costs, there are more and more relevant researches. This paper presents a use of camera sensors of method to implement real-time motorcycle detection and tracking. This method can be applied in a variety of road environment. For motorcycle detection, we use haar-like features as the digital image features of machine learning, and use adaboost and cascade classifier for training and detection, in order to achieve real-time detection purpose; For motorcycle tracking, we use the principle which the detection windows are neighboring and similar in the video. We conducted experiments in the different environments, such as urban streets of dynamic background, sunny and rainy city streets of static background. Our proposed method can effectively detect a variety of motorcycles, in sunny environment can achieve 92.9% precision rate and 94.2% recall rate, while in the high degree of complexity and the noise of the rainy environment can achieve 92.2% precision rate and 79.1% recall rate; And the proposed method of tracking in the urban streets also showed stable. The overall performance is 20 fps at 640×480 resolution.