Real-Time Driver Distraction Detection Using Lightweight Convolution Neural Network with Cheap Multi-scale Features Fusion Block

Degraded driving performance is closely related to distracted driving behavior. Therefore, in this paper we introduce a fast and accurate convolution neural network (CNN) architecture named CMFFNet, which is designed for the edge devices with limited computational resources and application scenario of detecting driver distraction. CMFFNet is built on a multi-branch architecture that uses CMFF blocks to guarantee efficiency. CMFF block is based on depth-wise separable convolution, which consists of a CMFF module, a multi-branch structure fuses filtered multi-scale spatial information to enhance feature representation capability, followed by a MAPGC bottleneck, utilizing two point-wise group convolutions and a channel shuffle operation to implement efficient channel projection. Compare with recently proposed lightweight CNNs, CMFFNet can achieve better performance on StateFarm’s Dataset, e.g. higher accuracy (absolute 5.5%) and lower latency (about 18 ms) compared with MobileNetV3-Large.