Automatic detection of vehicle occupancy and driver's seat belt status using deep learning

• Published in: Signal, image and video processing Vol. 17; no. 2; pp. 491 - 499
• Main Authors: Hosseini, Sara, Fathi, Abdolhossein
• Format: Journal Article
• Language: English
• Published: London Springer London 01.03.2023; Springer Nature B.V
• Subjects: Accuracy; Artificial neural networks; Automobiles; Computer Imaging; Computer Science; Deep learning; Image Processing and Computer Vision; Multimedia Information Systems; Original Paper; Pattern Recognition and Graphics; Seat belts; Signal,Image and Speech Processing; Transportation networks; Vision; Windshields; Car occupant detection; Deep learning; Seat belt status detection; Automated transport images analysis; Car windshield detection
• DOI: 10.1007/s11760-022-02244-w
• ISSN: 1863-1703

Increasing the number of personal cars on the transportation routes causes a heavy traffic load. In many countries, special lines for high occupancy vehicles have been developed to reduce the traffic load. Another issue in monitoring the transportation is the control of driving rules such as driver’s seat belts violation. In this paper, a new system based on deep learning models for detection of the occupants and the status of driver's seat belt is proposed. In this method, first, the windshield is detected using the YOLOv5s network. Then, the presence of passenger and driver’s seat belt rule violation is detected using deep learning-based models. To this end, the combinations of pre-trained the residual neural network with 34 layers (ResNet34) and power mean transformation layer along with spatial pyramid pooling or temporal pyramid pooling layers are employed. The proposed models were trained and evaluated on over 3500 images obtained from the Traffic Transport Organization. From the obtained results, the proposed model can detect the windshield with 99.7% accuracy. Also, the accuracy of the proposed models using independent random test set for occupant detection and drivers’ seat belt rule violations detection is 99.7% and 98.9%, respectively. The performance of the proposed method is better or comparable with state-of-the-art methods.