Highly Reliable Vehicle Detection through CNN with Attention Mechanism

Li-Wen Wang, Wan-Chi Siu, Xue-Fei Yang, Zhi-Song Liu, Daniel Pak-Kong Lun

Abstract: This paper presents a novel approach to provide reliable vehicle detection by using CNN-based approach and lane information. We firstly propose an adaptive RoI strategy that utilizes road lane information to give focus on frontal area for vehicle detection. Then we introduce a novel attention mechanism that automatically learns an attention map to refine the features for detection. Experimental results show a large improvement (+73 % on recall rate) for long-range (30 to 60m) vehicle detection which is extremely useful for users of driving assistant systems.