Original Article Proc IMechE Part D: J Automobile Engineering 1–16 Ó IMechE 2021 Article reuse guidelines: sagepub.com/journals-permissions DOI: 10.1177/09544070211025338 journals.sagepub.com/home/pid Deep learning based automatic vertical height adjustment of incorrectly fastened seat belts for driver and passenger safety in fleet vehicles Arif S xenol S xener 1 , Ibrahim Furkan Ince 2,3 , Husnu Baris Baydargil 3 , Ilhan Garip 4 and Oktay Ozturk 5 Abstract The recognition of incorrect fastening of seat belts is significant in passenger and driver safety for the automotive indus- try and public health. It should be made sure that the passenger’s seat belt is not only fastened but also correctly fas- tened across the body so that the passenger is adequately protected in the event of an accident. Current technology employs the buckle effect sensor, which merely solves the buckling detection problem, but there is no reliable solution for the correct positioning of the seat belt. Additionally, computer vision-based systems are still incapable of recognizing the incorrect positioning of seat belts when the training is performed by employing the subjects out of the fleet. Considering this fact, in this study, we propose a novel solution that employs a vision-based incorrect fastening seat belt detector to perform automatic vertical height adjustment independent from drivers and passengers for the fleet vehicles. We recognize the incorrect positioning of the seat belt inside the car by the acceptable distance of the seat belt from the neck of drivers or passengers to avoid neck injuries and the deaths caused by neck cuts. An extensive benchmarking is performed by comparing the three CNN architectures such as; DenseNet121, GoogLeNet (Inception-v3), ResNet50 with respect to sensitivity, specificity, precision, false-positive rate, false-negative rate, F1 score, and accuracy. Additionally, training and validation loss curves and accuracy curves are plotted for all the models. Later, the three mod- els are evaluated with a precision-recall (PR) curve at the end. According to the results, the DenseNet121 achieved the highest classification accuracy among the tested models with 99.95%. This paper includes information about the pro- posed system elements, registration of data, elaboration of data, program algorithm, testing the system in the lab, and on the vehicle. Keywords Vision-based automatic vertical height adjustment of seat belts, incorrectly fastened seat belts, passenger and driver safety, fleet vehicles, linear DC motor, convolutional neural network, deep learning Date received: 13 January 2021; accepted: 18 May 2021 Introduction Traffic accidents are among the most critical problems that cause injury, death, material loss, and psychologi- cal issues that need to be solved worldwide. 1–4 They are the leading cause of death, killing more than 1.2 million people annually worldwide, half of whom are between the ages of 15 and 44. 5 They cause significant disability in more than 50 million trauma patients. 6 The main innovations in road traffic safety that have helped reducing mortality are the correct use of seat belts. 7,8 It was stated that the seat belt could lessen the risk of severe injuries to the driver and front-seat passenger by 45% and judicious injuries by 50%. 9–12 The primary 1 Department of Mechanical Engineering, Nisantasi University, Istanbul, Turkey 2 Department of Digital Game Design, Nisantasi University, Istanbul, Turkey 3 Department of Electronics Engineering, Kyungsung University, Busan, Korea 4 Department of Electrical and Electronics Engineering, Nisantasi University, Istanbul, Turkey 5 Department of Computer Engineering, Nisantasi University, Istanbul, Turkey Corresponding author: Arif S xenol S xener, Department of Mechanical Engineering, Nisantasi University, Maslak Mahallesi, Tasyoncasi Sokak, No: 1V ve No:1Y Bina Kodu: 34481742 Sariyer, 34398 Istanbul, Turkey. Email: arif.sener@nisantasi.edu.tr