A two stage multi object tracking algorithm with transformer and attention mechanism.

Area of Science:

• Engineering Safety
• Computer Vision
• Artificial Intelligence

Background:

• Multi-object tracking faces challenges with occlusion and frequent identity switches (IDs) in engineering safety applications.
• Existing tracking-by-detection models struggle with accurate detection and maintaining target identity under challenging conditions.

Purpose of the Study:

• To propose an advanced multi-object tracking model integrating improved You Only Look Once Version 8 (YOLOv8) and High-Performance Multi-Object Tracking by Tracking Bytes (ByteTrack).
• To enhance object detection and re-identification capabilities to mitigate occlusion issues and reduce ID switches for improved tracking robustness.

Main Methods:

• Developed a novel Coordinate Attention Spatial Pyramid Pooling - Fast Conv (CASPPFC) module and integrated it with an improved Efficient Vision Transformer (EfficientViT) to enhance the YOLOv8 backbone.
• Introduced the Omni-Scale Network-Coordinate Attention (OSNet-CA) for effective Re-identification (Re-ID) feature extraction.
• Employed the Efficient Intersection over Union (EIoU) method in the association stage to refine target positional relationships.

Main Results:

• The improved model achieved 80.5% MOTA, 79.3% IDF1, and 64.2% HOTA on the MOT17 dataset.
• On the MOT20 dataset, the model reached 77.8% MOTA, 76.9% IDF1, and 62.4% HOTA.
• Demonstrated significant reduction in false positives, false negatives, and identity switches compared to baseline methods.

Conclusions:

• The proposed multi-object tracking model effectively addresses occlusion and ID switch challenges in engineering safety.
• The integration of advanced modules like CASPPFC, EfficientViT, OSNet-CA, and EIoU significantly boosts tracking accuracy and robustness.
• The model provides high-precision pedestrian tracking, crucial for timely detection of hazardous events and ensuring personnel safety in engineering environments.