Automated Traffic Surveillance Using Existing Cameras on Transit Buses

Area of Science:

• Transportation Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Traffic congestion significantly impacts daily commutes and requires accurate data for effective management.
• Current traffic detection methods using fixed or temporary detectors are limited by spatial and temporal sparsity.
• Previous research suggested using public transit buses as mobile surveillance platforms, validated through manual video analysis.

Purpose of the Study:

• To operationalize a vision-based vehicle counting methodology for practical traffic surveillance using existing bus sensors.
• To develop an automatic system for processing video imagery from transit buses to count vehicles.
• To leverage perception and localization sensors already equipped on buses for traffic data collection.

Main Methods:

• Utilized a state-of-the-art 2D deep learning model for frame-by-frame object detection.
• Employed the SORT (Simple Online and Realtime Tracking) method for tracking detected objects.
• Developed a counting logic to convert tracking data into vehicle counts and bird's-eye-view trajectories.

Main Results:

• Demonstrated the system's ability to detect and track vehicles in real-world video data from in-service buses.
• Successfully distinguished between moving vehicles and parked vehicles.
• Achieved high-accuracy bidirectional vehicle counts, validated through ablation studies and diverse weather conditions.

Conclusions:

• The proposed vision-based system effectively operationalizes transit buses for traffic surveillance.
• This automated method overcomes the limitations of traditional traffic detectors, providing more comprehensive traffic data.
• The system shows high accuracy and robustness, making it suitable for practical traffic management applications.