Deep Learning-based Vehicle Counting and Speed Estimation in Surveillance Videos using Object Detection

This paper introduces a Deep Learning-based Vehicle Counting and Speed Estimation (VCSE) system. The proposed VCSE system applies object detection techniques and custom-developed algorithms on surveillance videos. For object detection, the system uses the Single-Shot Detection algorithm and fine-tuned MobileNetV2 CNN over a custom dataset. Moreover, it uses a custom-developed algorithm for object tracking and speed estimation. After fine-tuning object detection model, the mAP score for object detection was 0.88. The parameters of the speed estimation algorithm were also fine-tuned using a curated calibration dataset that contains videos of the vehicles with varying direction, speed and environment. The accuracy of the speed estimation was around 98% after calibration. The proposed VCSE system finds its applications in law enforcement, smart-city planning and traffic management. It is a valuable tool for enhancing the safety and security posture of an organization by monitoring the vehicles’ speed on a real-time and 24/7 basis.