Object detection for ensuring car seatbelt buckling using the yolov7 algorithm

Abstract

Background: An estimated 3287 people die in accidents every day, and about 1.35 million people die in automobile accidents every year, and nearly half of all individuals involved in these accidents were not wearing their seat belts at the time of the crash. Car drivers who disregard safety regulations are the biggest contributors to most accidents, which may sometimes lead to death or severe injuries. This statistic exists even when research has shown that buckling a seat belt reduces fatalities and severe injuries by approximately 45% and 50%. Lack of safety precautions occurs despite the law in place and the numerous safety sensors and warning indicators embedded in modern cars (Advanced Driver Assistance System (ADAS) technology). Objectives: There is currently not a lot of research devoted to enforcing the buckling of a seat belt. Drivers and car occupants easily cheat the available seat belt technology for most cars, as they sometimes choose to buckle their car seat belts behind their backs to mute the seat belt alarm. Consequently, this dissertation presents an object detection model to monitor the buckling of a car seat belt between two (2) classes: buckled and unbuckled. The You Only Look version 7 (YOLOv7) object detection model was adopted because of its accuracy, robustness, and real-time capabilities. The main objectives of the study were to: 1. Develop and train an object detection model for a car seat belt using a graphics processing unit (GPU). 2. Deploy the training weights from a GPU platform to a Jetson Nano board for inferencing. 3. Test the deployed model using the NVIDIA DEEP STREAM object detection applications. Design: The data was collected by recording a video of a driver and vehicle when driving in different light conditions with different apparel. Some of the data was taken from friends, and some was downloaded online. The videos were converted to images and processed so that they could be ready for annotation, training, testing, and validation of the model. A model’s training was performed on a core i5 Windows computer processor (10th generation) with a GTX 1650 Nvidia GPU. The trained model weights were converted to lighter weights called tensorrt (this was done to increase the speed and optimise the deployed model) to accommodate the Nvidia Jetson Nano platform. An inference was implemented on the Jetson Nano board in a moving vehicle to examine the real-time capabilities of the model. Results: The present study achieved a mean average precision (mAP) of 99.6% when the threshold was set at 0.50. Furthermore, the model demonstrated an F1-score of 98%, along with precision rates of 99.7% and 99.5% for the buckled and unbuckled classes, respectively. Conclusion: This dissertation demonstrated YOLO's potential for car seat belt buckling detection. The results showed high accuracy even under challenging conditions like light variations, partial occlusion, blurry images, and occupants wearing seat belt-like clothing. The study also showed that GPU model training saves time. The model's successful deployment on a resource-efficient platform like the Jetson Nano development board allowed for its integration into modern car systems.