Passive Camera-Based Vehicle Orientation Estimation for Autonomous Systems

A critical task for autonomous vehicles is not only detecting surrounding objects but also predicting their orientation and heading, particularly for nearby vehicles. Understanding the direction a neighboring vehicle is facing and its potential trajectory enables autonomous systems to make informed navigation decisions, avoid collisions, and operate safely in traffic. This paper develops and evaluates two passive-sensing methods that leverage machine learning to predict vehicle orientation from two-dimensional (2D) images. The first approach employs deep learning to classify a vehicle's orientation into one of eight general directions. The second method utilizes a cascaded approach with object detection, bounding box area analysis, and regression to predict a more precise orientation of the vehicle. A dataset of 1,424 labeled images, each annotated with the relative heading difference of the distant vehicle with respect to the observing vehicle, was collected and used in both approaches. The findings of this research indicate that the second, cascaded approach is particularly effective, achieving a Mean Absolute Error of 5.06 degrees, demonstrating its potential for robust vehicle tracking applications.