Vehicle Detection and Tracking Techniques Using Yolo Detection and Kalman Filter Optimization

Abstract

This study develops a real-time vehicle detection, tracking, and counting system by integrating the You Only Look Once (YOLO) algorithm with the Kalman Filter. The system is designed to detect multiple types of vehicles, track their movements across frames, and count the number of vehicles crossing a detection line in traffic videos. YOLO is employed for vehicle detection by generating bounding boxes and class labels, while the Kalman Filter ensures robust and stable tracking, even in scenarios with interrupted detections or occlusions. The system was evaluated on real-world traffic videos and demonstrated high accuracy in vehicle detection, achieving a Mean Average Precision (mAP) of 96.7% after 200 training epochs. The system effectively counts vehicles with minimal errors, particularly for larger vehicles like cars and trucks. With an average processing speed of 25 frames per second (FPS), the system is capable of meeting real-time application demands. Additionally, dense traffic conditions with significant vehicle overlaps can impact detection and tracking accuracy. To address these limitations, future work is suggested, including retraining the model using localized datasets to improve performance under specific conditions, integrating complementary sensors such as LIDAR for better spatial awareness, and optimizing the system for deployment on embedded devices. This system has substantial potential for implementation in intelligent transportation applications, such as traffic flow analysis, violation detection, and data-driven transportation management, contributing to advancements in smart city infrastructure.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint