Occluded Vehicle Detection via Multi-Scale Hybrid Attention Mechanism in the Road Scene

Abstract

The obstruction of vehicles by surrounding vehicles, obstacles, etc. is a common phenomenon in the practical application of automatic driving. In view of the problem that the vehicle’s vision is affected by the occlusion, the vehicle feature information is incomplete, resulting in the low detection accuracy of the occlusion vehicle, and the occlusion vehicle detection method based on the multi-scale hybrid attention mechanism is proposed. The paper aims to fully excavate the advantages of multi-scale feature extraction, channel/space attention and other modules, and to design a multi-scale hybrid attention module suitable for occlusion vehicle detection to improve the detection accuracy of occlusion vehicles. Multi-scale features are enriched by the grouping convolution of different sizes of multi-scale feature extraction networks, and the parallel connection channels and spatial attention modules form different scale hybrid domain attention modules, which enhance the local feature information of the occluded vehicles and realize the reinforcement learning of multi-scale features and the suppression of occlusion interference information. Experimental results show that in the self-made occlusion vehicle dataset and the BDD100K occlusion vehicle dataset, the average mean accuracy of this method is 95.2% and 59.3%, respectively, which is 1.5% and 2.9% higher than that of the baseline network YOLOv5, respectively.