Research On Road Scene Understanding In Infrared Images

Road scene understanding plays a very important role in vision system of intelligent vehicle. Present research on intelligent vehicle system mainly focus on visible(RGB)images. However, the visible light camera is quite sensitive to illumination and shadow,even cannot work well during night-time driving condition. To overcome the limitations,we take advantage of infrared images. But it exists weakness that infrared images contain less texture features and boundaries are less legible. Therefore, we study the problems of road boundary and pedestrian detection in infrared images. The main work and achievement of this essay:1. For solving the problem of road boundary detection in infrared images, we propose a set of novel neighboring window features for describing infrared image patterns.Then, we use random forests to learn an infrared edge classifier, by which a probability road edge map can be generated. Finally, we obtain exact evaluation of road boundaries by means of boundary extraction based on the principle of Freeman chain code and second order polynomial approximation. Experiments conducted on many kinds of road images captured by real vehicle validate that the proposed approach can perform well with effectiveness.2. For solving the problem of pedestrian detection on infrared images, saliency mechanism of human visual system is integrated into pedestrian detection approach. Instead of traditional force searching strategy, we take the object edge information to generate the latent pedestrian candidate proposals, which suppresses the interference affected by the background on foreground object detection. With two cascade classifiers to make a further confirm of suspected area, we fulfill the task of pedestrian detection in infrared images. This essay conducts performance testing on our own infrared dataset and publicly available dataset. Results validate the effectiveness of our approach.