Anticipating Driving Behavior through Deep Learning-Based Policy Prediction

TL;DR

This paper presents a deep learning system that predicts driving behavior by integrating visual and depth data, achieving 50-80% accuracy and outperforming models using only video frames.

Contribution

The study introduces a novel multi-modal deep learning approach combining visual and depth features for accurate driving behavior prediction.

Findings

Integrated features improve prediction accuracy over video-only models.

The system achieves 50-80% accuracy in real-world driving scenarios.

Fusion of visual and depth data enhances model reliability.

Abstract

In this endeavor, we developed a comprehensive system that processes integrated visual features derived from video frames captured by a regular camera, along with depth details obtained from a point cloud scanner. This system is designed to anticipate driving actions, encompassing both vehicle speed and steering angle. To ensure its reliability, we conducted assessments where we juxtaposed the projected outcomes with the established norms adhered to by skilled real-world drivers. Our evaluation outcomes indicate that the forecasts achieve a noteworthy level of accuracy in a minimum of half the test scenarios (ranging around 50-80%, contingent on the specific model). Notably, the utilization of amalgamated features yielded superior performance in comparison to using video frames in isolation, as demonstrated by most of the cases.