Lightweight bus driver smoking detection

The smoking behavior of bus drivers poses significant risks to the operational safety and service quality of public transport systems. To achieve high precision and lightweight detection of bus driver smoking, this paper presents a novel method that integrates YOLOv7-tiny with a channel pruning algorithm. Firstly, a small target detection layer is introduced atop the original three detection layers to enhance the model's capability in detecting cigarettes in small targets. Secondly, the standard 3×3 convolution is replaced by the Depth Separable Convolution (DSConv), thereby substituting complex computations with simpler linear computations, consequently reducing the computational load of the model. Thirdly, a combination of the Complete Intersection over Union (CIoU) and the Normalized Wasserstein Distance (NWD) loss functions is employed. This not only optimizes the convergence speed of network training but also enhances the accuracy of cigarette detection in bus drivers. Additionally, the channel pruning algorithm is applied to further lighten the model while maintaining or improving detection accuracy. Finally, ablation experiments are conducted on a self-constructed dataset of bus driver smoking to validate the effectiveness of the proposed method. Experimental results demonstrate that the improved model achieves a 3.3% enhancement in detection accuracy compared to the original YOLOv7-tiny algorithm, with a reduction of 73.3% in parameters and 66.9% in computation. These outcomes underscore the feasibility of deploying the model on mobile devices.