Real-Time Speed Estimation on Indian Vehicles using Deep Learning and Computer Vision

Road commuting often bears unnoticed risks, contributing significantly to fatalities and injuries across the world. Alarmingly, India tops the list of harm caused by road accidents with 1.5 million annual road fatalities. In this scenario, real-time speed estimation, facilitated by vehicle-mounted cameras, can potentially be an excellent tool for averting accidents and enhancing traffic management. This paper explores the complexities of real-time speed estimation, emphasizing efficient object detection and tracking for vehicles on Indian roads through transfer learning over the base model of YOLOv8. The new model yielded an accuracy of 83 % on the Indian Vehicle Dataset. The proposed algorithm also addresses the challenge of relative speed estimation of other vehicles on the road from a vehicle mounted camera by observing the changing pixel areas of the vehicles on the video feed. Through this approach, we were able to estimate the change in the speed of vehicles satisfactorily, laying the groundwork for future advancements in practical realtime speed estimation.