Edge-Optimized Deep Learning for Early Driver Fatigue Detection in Trucking: A Spatiotemporal CNN–LSTM Approach

This study develops a lightweight vision-based system integrating deep convolutional and recurrent neural networks to identify early drowsiness in commercial heavy-duty operators in real time. Given that fatigue-related collisions account for approximately 13% of severe crashes in the freight sector, I evaluate whether a spatiotemporal model deployed on the vehicle can surpass established blink-monitoring methods—specifically PERCLOS—in both speed and reliability of the fatigue-alert signal. I trained MobileNet-LSTM architecture using the publicly available UTA Real-Life Drowsiness Dataset (RLDD; 30 hours of diversity-rich video from 60 professional drivers), in which naturally occurring facial fatigue patterns are labeled as alert, low-vigilance, or drowsy. The model processes a live camera feed to extract and encode the temporal cues (eye closures, yawns, head nods) that precede microsleep episodes. I evaluated detection latency and reliability against a benchmarked blink/PERCLOS threshold using a cross-validated laboratory protocol and an 8-week commercial on-road trial involving 12 operators. The deep architecture produced fatigue warnings with a median latency of 1.0 second, compared to 2.5 seconds for the PERCLOS algorithm (p < 0.001), alongside a reduction in the mean rate of false positives (0.4 against 1.2 per hour; p < 0.01), while the overall classification accuracy improved from 65% to 80%. Error rates remained stable across dynamic illumination environments and diverse driver demographics, with no statistically significant accuracy bias observed across driver subgroups, indicating robust generalizability within a heterogeneous population. Real-time inference on an automotive-class ARM processor operates at approximately 10 milliseconds per frame, reflecting an optimized convolutional neural network–long short-term memory architecture compressed to 1 megabyte. The results indicate that an edge-based deep learning paradigm capitalizing on spatiotemporal facial dynamics identifies driver fatigue with greater priority and reliability than conventional blink-count metrics. This research thereby presents a technically feasible, resource-constrained driver-monitoring platform that autonomously notifies drowsy commercial-vehicle operators, thus potentially averting collisions and satisfying forthcoming regulatory mandates that require in-cab drowsiness assessment systems.