A Study on Bus Passenger Boarding and Alighting Detection and Recognition Based on Video Images and YOLO Algorithm

Area of Science:

• Computer Vision
• Artificial Intelligence
• Transportation Engineering

Background:

• Accurate passenger origin-destination (OD) data is vital for intelligent public transportation and smart city development.
• Traditional data collection methods like manual surveys and smart card data have significant limitations in accuracy and completeness.
• Existing object detection algorithms struggle with the challenges of high-density, occluded, and scale-variant passengers in bus environments.

Purpose of the Study:

• To develop an enhanced object detection model for accurately identifying bus passenger boarding and alighting events.
• To improve the performance of the YOLO algorithm in complex on-vehicle scenarios with occlusions and varying passenger scales.
• To provide a robust solution for real-time passenger flow data collection to support intelligent transportation systems.

Main Methods:

• The study proposes an improved YOLOv8n model incorporating a DAC2f structure (deformable attention + C2f) for better feature extraction and background suppression.
• A SWD-PAN module was introduced for effective bidirectional cross-scale feature interaction to handle scale variations.
• WIoUv3 was employed to optimize sample weighting, particularly for small targets and non-standard passenger postures.
• The enhanced YOLOv8 model was integrated with DeepSORT for improved multi-object tracking stability.

Main Results:

• The improved YOLOv8 model achieved significant enhancements in precision (+3.68%), recall (+5.12%), and mAP (+6.26%) compared to the baseline.
• The model meets real-time processing requirements for bus passenger detection.
• Integration with DeepSORT resulted in a MOTA of 31.24% (2.6% higher than YOLOv8n) and MOTP of 88.06%, effectively reducing trajectory breakage and ID switching.

Conclusions:

• The proposed enhanced YOLOv8 algorithm effectively addresses the limitations of traditional OD data collection methods in public transportation.
• This research provides a robust technical foundation for the refined management of intelligent public transportation and the optimization of smart city transportation.
• The developed system offers a viable solution for real-time, accurate passenger flow analysis, contributing to smarter urban mobility.