Intelligent Vehicle Detection and Velocity Prediction with Deep Neural Networks

This paper proposes a highly efficient approach for estimating vehicle speeds using traffic surveillance video. The method enhances an existing state-of-the-art system that identifies 3D vehicle bounding boxes through vanishing point geometry combined with an object detection network. By substituting the original Retina Net detector with the more advanced YOLOv6 model, we achieve substantial improvements in processing speed while preserving the accuracy of speed estimation. Multiple versions of the proposed system are assessed based on vehicle detection performance and speed calculation accuracy. Comprehensive experiments conducted on different hardware platforms, including edge devices, reveal a marked increase in processing frame rate (FPS) compared to the previous state-of-the-art solution, with equal or better estimation precision. We further examine the balance between computational complexity and performance, demonstrating that compact models enhanced with post-training quantization provide the most effective solution for practical deployment scenarios. Using the Brno Comp Speed dataset, the optimized model surpasses earlier benchmarks in speed estimation accuracy, detection precision, and recall, while operating at nearly eight times higher speed. Additional testing on a separate dataset confirms the robustness and adaptability of the proposed approach across varying environments.