Abstract
Cyclist detection is an important task for ADAS System. Not only because cyclists are as vulnerable as pedestrians but also they share the same road with other vehicles with a comparable velocity, which makes them much easier to get involved in accidents and to suffer from severe injuries and even fatalities. The appearance of cyclists in the image obviously varies according to the viewpoints. This requires extra algorithms to deal with multi-view detection and the precision should be maintained as well.
An image based approach for the detection of cyclists in urban situation under different conditions is presented and evaluated. Compared to other works, the proposed study does not use the geometrical analysis of contours like Hough transformations, which especially used to detect the bicycle wheels. In our approach the possibility to distinguish between cyclists and pedestrians is focused only on the detection of the typical pedal movement of a cyclist in different positions. Our approach is based on different levels of processing. First, cyclist features are extracted from the pedal movement. The vectors flows are filtered using a probabilistic map which gives us a stable pattern of the pedal movement. In the second step, significant features are selected and tracked only inside the pedal movement region. The generated hypotheses are analysed in the spatial and temporal domains. A spectrum based classification is used in order to classify people as cyclist.
Results is presented and discussed. The proposed approach is tested and evaluated using a dataset of recorded image sequences.
KEYWORDS Cyclist detection, maps generation, HOG, Cosine Transfom, intelligent camera