Stereo Vision-Based Vehicle Distance Estimation with a Two-Stage Deep Learning Approach

This study presents a two-stage deep learning pipeline for vehicle distance estimation in urban environments, using synthetic stereo dataset generated via Unreal Engine 5.4. The Unreal Engine 5.4 is simulation tool is highly realistic and closely mimics real-world conditions, while being significantly cost-effective and more time-efficient to produce compared to collecting labeled data manually in real-world conditions. A drone-mounted stereo camera system with an 80 cm baseline captured 7,000 stereo image pairs across diverse vehicle positions and dynamic lighting conditions, annotated with precise Euclidean drone-to-car distances, automatically computed via Unreal Engine’s Physics Engine, ranging from 10 to 200 meters. The f irst stage uses YOLOv11 for vehicle detection, and it was trained on this synthetic dataset to identify vehicles in urban scenes. The second stage uses a Convolutional Neural Network (CNN) that processes stereo image crops and normalized bounding box dimensions to predict drone-car distance. The proposed approach uses geometric features (such as bounding box scaling) with learned visual features to enhance distance estimation accuracy. The pipeline achieves a 1.02 meters training RMSE and 4.2 meters validation RMSE, with a mean relative error of 4.02%. The system demonstrates a validation error of 2.2% relative to the maximum distance of 200 meters, demonstrating the effectiveness of the proposed approach within a synthetic environment. Real-world deployment remains a clear objective for future work. However, a key limitation lies in the domain adaptation gap, as our pipeline is trained solely on synthetic data may face challenges when directly deployed in real-world scenarios.