Enhancing Road Safety: Deep Learning-Based Automated Seat Belt Detection in Indian Traffic

The global surge in driving rates has led to a rise in accidents and injuries, particularly affecting low and middle-income countries. Seat belts play a crucial role in protecting passengers during accidents, yet current seat belt fastness detection systems in vehicles are often easily deceived. Over 1.2 million people worldwide die annually in traffic accidents, with more than 50 million sustaining injuries. In India alone, over 4 lakh accidents occur each year, resulting in at least 1.5 lakh fatalities. To address these alarming statistics, efforts have focused on studying seat belt detection to mitigate injury severity in vehicle crashes. Traditional algorithms, including Hough and Canny edge detection, based on driver visual monitoring to recognize seat belt usage often yield false positives due to varying vehicle types and backgrounds. To overcome this, Convolutional Neural Networks (CNNs) have been employed as a deep learning methodology for seat belt usage detection. The CNN approach involves constructing an initial ConvNet, trained using a Seat Belt dataset. Evaluation using the YawDD dataset with Inception V3 yielded a 56% accuracy, while ResNet50 and VGG16 demonstrated bias and inadequate learning for the no seat belt class. Notably, ResNet50, trained over the YawDD dataset for 50 epochs, outperformed other models with a 72% accuracy using transfer learning.