Efficient Pedestrian Detection in Urban Traffic Scenes

Pedestrians are important participants in urban traffic environments, and thus act as an interesting category of objects for autonomous cars. Automatic pedestrian detection is an essential task for protecting pedestrians from collision. In this thesis, we investigate and develop novel approaches by interpreting spatial and temporal characteristics of pedestrians, in three different aspects: shape, cognition and motion. The special up-right human body shape, especially the geometry of the head and shoulder area, is the most discriminative characteristic for pedestrians from other object categories. Inspired by the success of Haar-like features for detecting human faces, which also exhibit a uniform shape structure, we propose to design particular Haar-like features for pedestrians. Tailored to a pre-defined statistical pedestrian shape model, Haar-like templates with multiple modalities are designed to describe local difference of the shape structure. Cognition theories aim to explain how human visual systems process input visual signals in an accurate and fast way. By emulating the center-surround mechanism in human visual systems, we design multi-channel, multi-direction and multi-scale contrast features, and boost them to respond to the appearance of pedestrians. In this way, our detector is considered as a top-down saliency system. In the last part of this thesis, we exploit the temporal characteristics for moving pedestrians and then employ motion information for feature design, as well as for regions of interest (ROIs) selection. Motion segmentation on optical flow fields enables us to select those blobs most probably containing moving pedestrians; a combination of Histogram of Oriented Gradients (HOG) and motion self difference features further enables robust detection. We test our three approaches on image and video data captured in urban traffic scenes, which are rather challenging due to dynamic and complex backgrounds. The achieved results demonstrate that our approaches reach and surpass state-of-the-art performance, and can also be employed for other applications, such as indoor robotics or public surveillance.