On-Road Vehicle Detection under Varying Lighting Conditions

摘要

On-road vehicle detection round the clock is important for driver assistance systems and autonomous driving. However, varying lighting conditions is a big challenge for outdoor scenarios. In this paper, a robust vehicle detection system under various lighting conditions is developed: First, different CNN(Convolutional Neural Network) models are adopted for different illuminations and they are merged into a single network to save the computational resources. Second, a target-oriented scene classification module is proposed to co-adapt scene classifiers with vehicle detectors. The scene classifier is explicitly optimized towards a better detection performance. Experimental results based on a database with varying lighting conditions show the potential of the developed method. The developed framework can also be easily extended to other weather conditions, such as rainy, snowy, foggy days, and so on.