Reduced CPU Workload for Human Pose Detection with the Aid of a Low-Resolution Infrared Array Sensor on Embedded...

Area of Science:

• Computer Vision
• Embedded Systems
• Machine Learning

Background:

• Modern embedded systems offer significant processing power for edge computing and computer vision tasks like pose detection.
• Existing frameworks (e.g., MediaPipe) consume high CPU resources, leading to wasted power and unnecessary data generation, even without active subjects.
• False detections in current systems further exacerbate inefficiency and resource drain.

Purpose of the Study:

• To develop an efficient hybrid computer vision system for embedded applications.
• To reduce the CPU workload and power consumption of pose detection algorithms.
• To minimize false detections in low-activity environments.

Main Methods:

• Utilized a low-cost infrared thermal sensor array to trigger pose detection.
• Developed a lightweight algorithm for person detection and isolation in thermal images.
• Integrated the thermal detection system with MediaPipe's pose detection on single-board computers.

Main Results:

• Achieved significant reduction in average CPU workload, particularly in low-activity scenarios.
• Effectively eliminated MediaPipe's false detections by controlling execution based on thermal input.
• Demonstrated up to 30% power saving in optimal conditions.

Conclusions:

• A hybrid approach using thermal sensing to control computer vision algorithms is highly effective for embedded systems.
• This method optimizes resource utilization, reduces power consumption, and enhances system efficiency.
• The system offers a practical solution for energy-efficient monitoring and security applications.