Streamlit Application for Helmet Detection Based on YOLOS: Case Study Indonesia

Traffic violations are still the biggest problem in the development of traffic in Indonesia. The three biggest violations are (1) running a red light, accounting for 42% of violations, (2) not wearing a helmet, accounting for 23% of violations, and (3) violating traffic signs, accounting for 9% of violations. The violation of not wearing a helmet is the second largest violation in Indonesia, which is due to the lack of awareness of safety while driving. Additionally, the supervision carried out by the security forces is also somewhat inadequate. Supervision is only conducted in certain areas, leading to a high number of violations. To address these issues, the YOLOS (You Only Look at Once Sequence) model, using Vision Transformer and Pytorch, was developed with the assistance of the AI City Challenge 2023 dataset. This model, trained for 200 epochs, is capable of detecting riders who wear helmets and those who do not. It achieves average precision and average recall values for objects with large area coverage in the precision and recall curves (large area), namely 38.8% and 57.2%, and on average precision with an IoU parameter of 0.5, obtaining a value of 63.5%. The implementation of the model in the Streamlit framework to create a system that can be used for detection was successfully completed. The system is capable of performing detection on a single image, webcam video, and video from RTSP CCTV.