Nighttime Vehicle Detection Based on Lane Information and Detector with Convolutional Neural Network Features

In this paper, a robust nighttime vehicle detection method based on the image sequence for Intelligent Headlight Control system is proposed, which could recognize vehicles accurately and efficiently at night. Previous researches recognizing the vehicle light from nighttime images can be divided into three steps: light blobs detection, classification based on appearance or movement of vehicle, post-processing. Unlike previous work in the area, which used one analyzed threshold or improved Laplacian of Gaussian operator in the whole image, we implement Progressive Probabilistic Hough Transform to detect lanes at night, which restrict the region of interest firstly. And then a simple light blob detector is used to obtain the candidate blobs, which promotes the robustness of the algorithm. Furthermore, CNN instead of traditional machine learning algorithm is applied again to distinguish the nuisance light from truly vehicle lights in candidate blobs on account of different frames. After some effective post-processing, we finally adopt the method to some nighttime image sequences. The experimental results demonstrate the satisfied performance of the propose algorithm maintaining reasonable detection speed and accuracy.