A Comparison of Machine Learning and Deep Learning Techniques for Activity Recognition using Mobile Devices

摘要: We have compared the performance of different machine learning techniques for human activity recognition. Experiments were made using a benchmark dataset where each subject wore device in pocket and another on wrist. The comprises thirteen activities, including physical common postures, working activities leisure activities. apply methodology known as recognition chain, sequence steps involving preprocessing, segmentation, feature extraction classification traditional methods; we also tested convolutional deep networks that operate raw data instead computed features. Results show combination two sensors does not necessarily result an improved accuracy. determined best results are obtained by extremely randomized trees approach, operating precomputed features from wrist sensor. Deep architectures did produce competitive with architecture.