A Novel IoT-Based Smart Recognition System for Distracted Drivers Using ACS Optimized Deep CNN Models

Abstract

Human drivers display various driving styles and experiences due to their distinct driving characteristics. According to worldwide statistics, human error is the most common cause of car accidents. A catastrophic and tragic accident might occur due to a human driver’s little distraction; hence, the driver must remain aware, especially when high-speed travel is permitted. Conventional methodology and approaches for detecting distracted driving behavior cannot capture complex temporal characteristics of driving actions. However, the advancement of deep learning algorithms and the Internet of Things (IoT) has made it possible to predict, detect, and analyze human driver–distracted behavior more efficiently and effectively. Intending to improve transportation safety and reduce road accidents, in this paper, we contribute by proposing an IoT-based framework for monitoring and detecting distracted driver behavior in smart cities. Moreover, we have proposed a hybrid approach for the detection of distracted human driver behavior by aggregating handcrafted and deep CNN features. In our proposed technique, we leverage four deep CNN models for extracting deep features, and then, handcrafted features are extracted using histogram of oriented gradient (HOG). Afterward, ant colony system (ACS) optimization is used as a feature selection technique. It takes the extracted features, removes the redundant and irrelevant information, and improves the classification performance. After feature selection, the selected features are fused before passing them to different variants of support vector machine (SVM) and K nearest neighbor (KNN) classifiers, and their performances are evaluated using standard measures. Performance comparison of deep models and the proposed model is carried out. It is observed through extensive experimentation that amongst the deep models, Inception V3 shows the shortest training time of 338.9 s on 1000 selected features using fine KNN, and the obtained accuracy is 94.21%. The best accuracy and training time of 95.26% and 987 s are achieved using the fine KNN classifier on 1000 selected hybrid features, respectively. Moreover, it was observed that with the increase in the number of selected features, the performance of the proposed model is increased in terms of accuracy.