Effective of contrastive learning framework in driver behavior analysis

The demand for advanced driver behavior analysis systems to support the car driver arises, leading to reduced accidents. The solutions have been researched and developed for a long time, but the results have recently been acknowledged since some deep learning methods have been published widely. Our paper proposes some modifications and one of the effective deep learning models, Contrastive Learning Framework (CLF), for better understanding and impact overall. However, it met a lot of challenges such as data imbalance and real-time predicting problems. In more detail, we propose the CENCE loss function for computing comparable visual features both positive and negative, and Cross Stage Partial Technique (CSPNet and CSPResnet) to improve the outcome in the base encoder. Our approach is evaluated on published datasets and the obtained results represent some positive performance in the analysis of driver behavior.