Abstract

Intelligent transportation systems require automatic traffic flow prediction and vehicle speed estimation from images or videos. However significant occlusions, variations in vehicle appearance, and differing camera perspectives complicate these tasks. Addressing these challenges requires robust algorithms for vehicle detection and tracking that can adapt to changes in car orientation and illumination. In this work, we propose a novel approach that combines modern deep learning models with classical computer vision techniques to enhance both detection and tracking accuracy. We begin by improving the baseline performance of Faster Region-based Convolutional Neural Network (R-CNN) through strategic modifications, leading to reduced computational complexity and enhanced detection accuracy. To further improve tracking, we integrate Linear Constrained Tracking (LCT) with the Deep Simple Online and Realtime Tracking (DeepSORT) algorithm, effectively minimizing false positives caused by unreliable detections. Additionally, we address the challenge of transitioning from image space to the real world by estimating rectifying transformations using vanishing points, which significantly improves vehicle speed estimation. The results of our experiments show that our suggested Improved R-CNN_DeepSORT framework improves both accuracy and performance by a large amount, beating out other methods in the field.