Forward-looking omnidirectional infrared pedestrian detection for driver assistance

To improve the intelligent driving abilities about the day-night situation awareness and large airspace real-time acquisition, we employ a forward-looking omnidirectional infrared (FLOIR) system and propose a FLOIR pedestrian detection strategy which consists of off-line training and on-line identification. Firstly, the basic scheme of proposed strategy is given generally, and the thoughts of off-line training and on-line identification are introduced briefly. Secondly, the principle of off-line training is illustrated in detail, in which the EHOG describer is created to extract the pedestrian feature and the FART neural network is modified to train the samples. Thirdly, the on-line identification principle is described, integrated with the proposed methods of ROI segmentation, distortion correction, feature extraction and matching. Finally, to verify the adaptability of FLOIR pedestrian detection relative to the climates, the summer and winter experiments, the contrast experiment with deep learning method and the multi-scale pedestrian detection experiment are carried out. The results show that: the proposed strategy has better robustness and detection effect than the traditional methods of FART, ARTMAP and deep learning, and the all-weather accuracy of infrared pedestrian detection is more than 83%. In the next work, the infrared pedestrian database will be improved to further increase the detection accuracy.