In recent years, driver-assistance systems are very actively developed in cooperation with the industrial-government-academic complex to reduce the pedestrian deaths and injuries. Many of practical systems employ the RADAR, laser rangefinder, infrared cameras, or stereo cameras. However, it is important to develop a system using a single camera to reduce the costs for automobile markets cars. In this paper, we propose a method for pedestrian recognition near the zebra crossings by a single camera because many traffic accidents occur near the zebra crossings in Japan. Our algorithm extracts the pedestrian regions near the zebra crossings by separating from background and other objects. Our system consists of two main algorithms: a zebra crossing detection algorithm and a pedestrian recognition algorithm. First, the image captured by the camera mounted on a car is processed by these two algorithms. Then, the outputs of them are integrated to extract the pedestrian regions near the zebra crossing. In the zebra crossing detection algorithm, we utilize the cross ratio of the four collinear points. Because the cross ratio is preserved under the perspective projection, the cross ratio of the zebra crossing in 3D space is the same as that in its image. We first detect the edges of the input image to extract the boundaries of the white lines on the road. To enhance the robustness, we use the binarized image in addition to the edge image to detect the boundaries. Then, we classify vertical edges into two kinds: the boundaries where the intensity changes from dark to bright or from bright to dark. Finally, the cross ratio of four collinear points are calculated and verified if its value is within the range decided by the Japanese law. On the other hand, the pedestrian recognition algorithm is based on learning the database. We employ the cascaded AdaBoost classifier and Haar-like features proposed by Viola and Jones for training and recognition. The Haar-like features are extracted using a large number of simple appearance filters. It is suitable for detecting small details of pedestrian, such as hands and legs. The number of Haar-like features used in our system is 285223. AdaBoost is used to select a small set of features and to train the classifier. For each stage of the cascade, AdaBoost iteratively constructs a weighed linear combination of simple classifiers. Each classifier has a threshold of one feature value. The iteration of each stage of the cascade is stopped when a certain performance value is achieved. The training process is continued with the next stage ofthe cascade. Experimental results show that our system can detect the pedestrian regions near the crossings. Future works focus on the real-time computation of the proposed algorithms and the optimization of parameters of the training step.
Pedestrian recognition by a single camera for driver assistance
Fußgängererkennung mit einer einzelnen Kamera zur Fahrerassistenz
2008
8 Seiten, 8 Bilder, 1 Tabelle, 11 Quellen
Conference paper
English
Pedestrian recognition by a single camera for driver assistance
Automotive engineering | 2008
|Pedestrian Recognition by a Single Camera for Driver Assistance
British Library Conference Proceedings | 2008
|F2008-08-118 PEDESTRIAN RECOGNITION BY A SINGLE CAMERA FOR DRIVER ASSISTANCE
British Library Conference Proceedings | 2008
|Stationary video-based pedestrian recognition for driver assistance systems
Automotive engineering | 2009
|Component-based pedestrian recognition for advanced driver assistance systems
Tema Archive | 2012
|