Dual-Mode Serial Night Road Object Detection Model Based on Depthwise Separable and Self-Attention Mechanism

Existing road detection at night mainly focuses on vehicle detection, but pedestrian detection is also particularly important in the complex environment of night roads. Based on the idea that one module only focuses on one task, this article proposes a night object detection (OD) method under a dual-module framework, which relieves the pressure of the OD algorithm facing both prediction box and classification tasks. In order to perform a multiscale fusion of local features and global features and achieve rapid acquisition of vehicle and pedestrian prediction boxes, Module 1 employs a lightweight depthwise separable network. In Module 2, the network performance is enhanced by improving the internal structure of the residual block and adding the attention mechanism. Finally, we can obtain a higher useful feature identification rate and better classification accuracy. The efficiency of the proposed approach has been evaluated on two benchmark datasets, namely, the Berkeley deep drive (BDD) dataset and the Hong Kong Vehicle dataset. Experimental results demonstrate that the proposed method outperforms several state-of-the-art OD methods, and exhibits a superior accuracy rate of 92.46% under a high standard intersection over union (IoU) threshold while maintaining a high processing speed of 42.3 frames/s. Consequently, these results verified the potential of the proposed method for advancing the field of OD at night and it is expected to improve the safety of driverless technology in complex environments at night.