Multi-model Deep Learning Approaches for Vehicle Speed Estimation

Abstract

Accurate vehicle speed estimates are critical for traffic monitoring and management. Traditional speed estimation approaches, such as radar guns or pixel displacement methods, are resource-intensive and frequently struggle in real-time or dynamic traffic situations. This paper describes a combination of YOLO for real-time object identification and Long Short-Term Memory (LSTM) networks for vehicle speed prediction that overcomes these constraints. The ability of YOLO to recognize cars with high precision, along with the capability of LSTM in processing sequential data, makes this approach extremely good for calculating speed from video data. Using the VS13 dataset of 400 single-vehicle videos, YOLOV5 detects vehicles and generates bounding box areas for LSTM to estimate speed. The performance across 13 car models averaged 5.79 km/hr RMSE. The accuracy and generalizability of the model can be improved in the future by adding outside variables like weather, road conditions, and driver behavior.