Distraction Detection in Driving Using Pose Extraction and Machine Learning Techniques

There are a lot of technologies have been implemented in the modern cars these days, ranging from preventive braking system to driving assistance. It is believe that equipping car with modern technology would enhance safety. However, there are other aspects need to be considered in safety driving, that is human factor. Human factors are believed have high contribution in road accidents. Human factors such as emotional control and distraction may leads to dangerous driving behaviours. This study investigates the driver pose distraction. The distracted driver pose detected through pose detection using landmark feature extraction. The dataset used in this study is State Farm Distracted Driving Detection which is a secondary dataset. The extracted driver pose are in numeric form and used as the input for the classification model to detect ten kinds of driving distraction ranging from safety driving to talking to passenger. Four classifiers used to detect distracted driver are CatBoost, Decision Tress, Random Forest and SVM. The model performance evaluated using Precision, Recall, Accuracy, and F1-score. Furthermore, prediction time recorded to determine the best model to detect driver distraction. The results shows that CatBoost model outperformed all other model with F1-score and prediction time of 0.899 and 0.014 second. Random Forest model performance is the second highest compared to four other with F1-score and prediction time of 0.881 and 0.028 seconds. CatBoost model is the best model even though Cat Boost model execution time is not the highest and the model performance is not significantly different compared to Random Forest.