Boosting Multi-Vehicle Tracking with a Joint Object Detection and Viewpoint Estimation Sensor.

Area of Science:

• Computer Vision
• Intelligent Transportation Systems

Background:

• Multi-vehicle detection and tracking are crucial for traffic monitoring.
• Existing methods often lack precise trajectory estimation due to limited observational data.

Purpose of the Study:

• To develop an intelligent visual sensor for enhanced multi-vehicle tracking.
• To improve trajectory estimation by integrating vehicle pose information.
• To introduce a new benchmark dataset for evaluating tracking algorithms.

Main Methods:

• Adapted an Extended Kalman Filter (EKF) to utilize vehicle detections and estimated poses.
• Developed two methods for simultaneous object detection and pose estimation: a Histogram of Oriented Gradients (HOG) with Support Vector Machines (SVMs) detector, and a modified Faster R-CNN deep learning model.
• Released the GRAM Road Traffic Monitoring (GRAM-RTM) dataset with over 700 vehicles annotated across 40,300 frames.

Main Results:

• Integrating vehicle pose observations into the EKF significantly improves trajectory estimation accuracy.
• Both HOG-SVM and Faster R-CNN approaches demonstrated effectiveness in simultaneous detection and pose estimation.
• The GRAM-RTM dataset provides a comprehensive resource for evaluating multi-vehicle tracking algorithms.

Conclusions:

• Simultaneous integration of vehicle localization and pose estimation enhances tracking performance.
• The proposed methods and dataset contribute to advancing multi-vehicle tracking in traffic monitoring.
• The GRAM-RTM dataset is expected to become a standard benchmark for the research community.