Harnessing Deep Learning and CGANs for Accurate Detection of Abnormal Driver Behaviors Using Skeleton Keypoints

This study investigates the use of deep learning models for detecting abnormal driver behavior by analyzing skeleton keypoint data extracted from the Driving Monitoring Dataset (DMD) and a custom real-world dataset. The objective was to evaluate various models for their ability to classify driver actions, including common distractions such as reaching for objects and using a phone. In the first experiment, several models, including CNN, CNN-LSTM, and CNN-GRU, were tested on the DMD dataset, with the CNN achieving the highest classification accuracy of 96.2%. The second experiment extended the analysis to a real-world dataset, recorded under diverse environmental conditions, to evaluate the models’ generalization capabilities. The CNN-GRU model outperformed others in this scenario, with an accuracy of 94%, demonstrating its robustness in the face of dynamic conditions, such as occlusions and lighting variations. To address the challenge of limited data for rare behaviors, Conditional Generative Adversarial Networks (cGANs) were employed to augment the dataset, resulting in significant improvements in model performance for rare actions. The results highlight the importance of combining spatial and temporal data for robust driver behavior classification. The study also emphasizes the need for data augmentation techniques, like cGANs, to overcome the challenges of data scarcity in real-world scenarios. By synthesizing data for rare behaviors, the model’s ability to generalize improved significantly, e nabling b etter detection of behaviors that are otherwise underrepresented in training datasets. These findings c ontribute t o t he d evelopment o f more accurate and reliable driver behavior detection systems for real-world applications.