Deep Learning-based Vehicle Counting and Speed Estimation in Surveillance Videos using Object Detection
Year: 2025
Type: conference-paper
Abstract: This article proposes a deep learning-based system that utilizes object detection techniques for vehicle counting and speed estimation in surveillance videos. The Vehicle Counting and Speed Estimation (VCSE) system integrates the SSD (Single Shot MultiBox Detector) algorithm and the MobileNetV2 model for accurate and efficient vehicle detection. To improve object detection performance, we conducted fine-tuning on the SSD and MobileNet models using a dataset collected from the surveillance system. We have also developed object tracking and speed estimation algorithms that leverage the object detection output to count and estimate vehicle speeds. The speed estimation performance has improved by performing the calibration activity. It involves generating a dedicated dataset of calibration videos covering various vehicle speed variations. Through this process, we achieved a remarkable reduction in error close to 2%, significantly improving the system's accuracy. The integrated system demonstrates promising results, with object detection performance reaching mAP score of 0.88 and the speed estimation accuracy around 98%. The proposed VCSE system has profound implications for law enforcement, safety, and security, providing a valuable solution for monitoring speed limits, enhancing road safety, and strengthening overall security measures.
Institution: Raja Ramanna Centre for Advanced Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4410809836