Detecting Abnormal Driving Behavior Using Modified DenseNet

Abstract

Car accidents have serious consequences, including depletion of resources, harm to human health and well-being, and social problems. The three primary factors contributing to car accidents are driver error, external factors, and vehicle-related factors. The main objective of this paper is to address the issue of car accidents caused by driver error. To achieve this goal, a solution is proposed in the form of a modified version of the Dense model, called the 1Dimention-DenseNet, specifically designed to detect abnormal driving behavior. The model incorporates adapted dense blocks and transition layers, which enable it to accurately identify patterns indicative of abnormal driving behavior. This paper compares the performance of the 1D-DenseNet model to the original DenseNet model in detecting abnormal driving behavior in the Kaggle distracted driver behavior dataset. Results show that the 1D-DenseNet model outperforms the original DenseNet model in classification and validation accuracies, loss, and overhead. The 1D-DenseNet, after 100 epochs of training using Keras on top of TensorFlow, the 1D-DenseNet achieved a categorical cross-entropy loss of 0.19 on the validation set, with classification and validation accuracies of 99.80% and 99.96%, respectively. These findings demonstrate the effectiveness of the 1D-DenseNet model in improving the detection of abnormal driving behavior.

Keywords

• Car accidents
• Deep Learning
• DenseNet