HitoMi-Cam: A Shape-Agnostic Person Detection Method Using the Spectral Characteristics of Clothing

Area of Science:

• Computer Vision
• Edge Computing
• Sensor Fusion

Background:

• Convolutional Neural Network (CNN) object detection struggles with shape variations not present in training data.
• Existing methods lack robustness in unpredictable environments, limiting real-world applications.
• A need exists for shape-agnostic person detection systems capable of real-time edge deployment.

Purpose of the Study:

• To implement and evaluate a spectral-based person detection method, HitoMi-Cam, on physical hardware.
• To assess the practical viability of HitoMi-Cam on resource-constrained edge devices without a GPU.
• To demonstrate HitoMi-Cam's effectiveness as a complementary tool to CNNs in challenging scenarios.

Main Methods:

• Developed HitoMi-Cam, a lightweight, shape-agnostic person detection system utilizing clothing spectral reflectance properties.
• Implemented the system on an edge device, evaluating performance without GPU acceleration.
• Tested performance in simulated search and rescue scenarios and various evaluation conditions.

Main Results:

• Achieved a processing speed of 23.2 frames per second at 253 × 190 pixels, indicating real-time capability.
• HitoMi-Cam demonstrated a 93.5% average precision (AP) in simulated search and rescue, significantly outperforming CNNs (best AP 53.8%).
• Maintained minimal false positives across all evaluation scenarios.

Conclusions:

• Spectral-based person detection is a viable option for real-time operation on edge devices.
• HitoMi-Cam offers a robust, shape-agnostic solution for unpredictable environments like disaster rescue.
• The method serves as a valuable complementary tool to CNN-based detectors under specific conditions.