People Counting Using YOLO-Based Detection and Clustering for a Mobile Robot

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Accurate people counting is crucial for intelligent monitoring systems.
• Mobile robot applications present unique challenges like moving cameras and occlusions.
• Existing methods struggle with dynamic environments and limited computational resources.

Purpose of the Study:

• To develop an effective people counting algorithm for mobile robots operating in dynamic indoor environments.
• To address challenges posed by moving cameras, varying conditions, and occlusions.
• To create a computationally efficient method suitable for resource-constrained robotic platforms.

Main Methods:

• Utilizes the You Only Look Once (YOLO) detector for visual people detection and bounding box identification.
• Extracts features from detected individuals' regions of interest (ROIs).
• Employs an encoder for feature dimension reduction, followed by K-means or SK-means clustering to estimate the number of unique individuals.

Main Results:

• The algorithm was validated on a dataset of 19,350 RGB images from 45 sequences, including mobile robot scenarios.
• Achieved a Mean Absolute Error (MAE) of 1.11 using YOLOv10n with K-means or SK-means clustering.
• Demonstrated effectiveness in various configurations and significantly reduced clustering time.

Conclusions:

• The proposed people counting method is effective for mobile robot applications in dynamic indoor settings.
• The encoder-based approach enhances efficiency, making it suitable for resource-limited platforms like the Jetson Nano.
• This algorithm offers a viable solution for real-time crowd estimation in intelligent monitoring systems.