抄録
Pedestrian protection is an active area in the research field of advanced driver assistance systems. A pedestrian who intends to cross the road is more critical to the driver than one who has no intention. This paper proposes a method for estimating the crossing intention of a pedestrian on a sidewalk from the pedestrian's posture and change in posture in multiple frames during a short period of video images. It is assumed that a pedestrian in the input video is detected and tracked with some method, and the bounding boxes of a pedestrian in the input video frames are already available. First, the appearance feature of the pedestrian posture is extracted from the image and the feature vector is quantized, and a sequence of the appearance feature of the posture coded by vector quantization is obtained for a single pedestrian in multiple image frames. Then, the probability that the pedestrian intends to cross the street is calculated by a probabilistic model from the sequence of posture for the past n frames. We evaluate the method in an indoor simulated environment and demonstrate the results and performance.
