Benchmarking YOLOv5 and YOLOv7 models with DeepSORT for droplet tracking applications

Area of Science:

• Microfluidics
• Computer Vision
• Data Analysis

Background:

• Tracking droplets in microfluidic devices is crucial for analyzing physical quantities but presents significant challenges in tool selection.
• Existing object detection and tracking algorithms require customization for effective droplet identification and tracking in microfluidic videos.

Purpose of the Study:

• To benchmark the performance of You Only Look Once (YOLO) versions v5 and v7, combined with the Simple Online and Realtime Tracking with a Deep Association Metric (DeepSORT) algorithm, for microfluidic droplet tracking.
• To evaluate training and inference times across different hardware configurations for custom microfluidic droplet datasets.

Main Methods:

• Trained multiple YOLOv5 and YOLOv7 models alongside the DeepSORT network for droplet identification and tracking.
• Compared the performance of YOLOv5 and YOLOv7 with DeepSORT regarding training duration and video analysis speed on various hardware setups.
• Assessed the impact of DeepSORT on overall tracking speed and real-time performance.

Main Results:

• While YOLOv7 demonstrated a 10% speed improvement over YOLOv5, real-time tracking was primarily achieved by lighter YOLO models on an RTX 3070 Ti GPU.
• The DeepSORT algorithm introduced significant computational overhead, impacting overall tracking speed and limiting real-time capabilities with heavier YOLO models.
• Performance varied across different hardware configurations, highlighting the importance of hardware in achieving real-time microfluidic droplet analysis.

Conclusions:

• The choice of YOLO model (lighter versions) and hardware (e.g., RTX 3070 Ti GPU) is critical for achieving real-time microfluidic droplet tracking when using DeepSORT.
• This benchmark provides valuable insights for researchers selecting appropriate YOLO and DeepSORT configurations for their specific microfluidic applications and hardware constraints.