Physical Activity Recognition Based on Motion in Images Acquired by a Wearable Camera

Area of Science:

• Computer Vision
• Human Activity Recognition
• Wearable Technology

Background:

• Standard activity recognition often relies on direct visual input of the subject.
• Wearable cameras offer a unique perspective but present challenges in data interpretation.
• Extracting meaningful information from indirect visual cues is crucial for advanced human-computer interaction.

Purpose of the Study:

• To develop and evaluate a new technique for extracting and assessing physical activity patterns from wearable camera image sequences.
• To analyze physical activity indirectly through camera motion, without direct wearer imagery.
• To compare the proposed method's accuracy and efficiency against existing feature detection techniques.

Main Methods:

• A multiscale approach was used for pixel correspondence identification.
• Motion feature extraction was performed based on motion statistics in each frame.
• Physical activity was determined using global motion distribution and tested with K-Nearest Neighbor (KNN), Naive Bayesian, and Support Vector Machine (SVM) classifiers.

Main Results:

• The proposed pixel correspondence technique is more accurate and computationally efficient than Good Features and Speed-up Robust Feature (SURF) detectors.
• The method successfully recognized various physical activities from real-world video data.
• Different machine learning classifiers achieved similar performances when provided with specific motion features.

Conclusions:

• The developed technique effectively recognizes physical activities indirectly through wearable camera motion.
• The approach offers a computationally efficient and accurate alternative to standard activity recognition methods.
• The findings highlight the potential of motion-based analysis for activity recognition in wearable sensing applications.