Estimating the Speed of Nearby Vehicles with a Single Onboard Camera by Smooth Kernel Regression

Area of Science:

• Computer Vision
• Robotics
• Automotive Engineering

Background:

• Estimating vehicle speed is crucial for advanced driver-assistance systems (ADAS).
• Existing methods often rely on expensive or complex sensor suites.
• A real-time, camera-only solution is needed for cost-effectiveness and broader adoption.

Purpose of the Study:

• To develop and evaluate a real-time, camera-only pipeline for estimating the relative speed of nearby vehicles.
• To provide a low-cost alternative or complement to active sensors like LiDAR or radar.
• To assess the accuracy and limitations of the proposed method.

Main Methods:

• Vehicle detection and tracking using a Convolutional Neural Network (CNN), specifically YOLOv8.
• Distance estimation based on bounding box angular width, class-dependent priors, and camera intrinsics.
• Nadaraya-Watson kernel smoothing applied to the distance sequence to estimate relative speed via analytic differentiation.

Main Results:

• The system successfully estimates vehicle distance and relative speed using only monocular camera input.
• Evaluation on synthetic data revealed a positional bias away from the image center.
• A trade-off between stability and lag was observed due to the smoothing kernel.

Conclusions:

• The proposed detector-agnostic pipeline effectively estimates vehicle speed from monocular camera data.
• This approach offers a viable, low-cost solution for driver assistance applications.
• Further validation in real-world scenarios is necessary to address identified limitations.