Human activity recognition using CNN

Abstract

In recent years, human movement analysis in real-life settings outside the laboratory has experienced increasing attention in sports and medical applications. At the same time, wearable sensors have been evolved as valuable tools. This has allowed to acquire large-scale human movement data that are typically complex, heterogeneous, and noisy. In this context, modelling approaches are required, as the measured quantities often do not directly reflect meaningful biomechanical variables. More recently, machine learning methods have emerged as promising modelling tools that exploit unstructured data for estimating relevant target variables, such as joint kinematics and dynamics. Although research in this field is still going on, there is a great potential not only to enlarge the range of numerous applications in sports but also to obtain biomechanical measures used to infer the load on body structures, such as joint forces. This applies in particular to unique sports such as ice hockey skating. Furthermore, little research has been conducted with respect to the direct estimation of biomechanical surrogate measures for knee joint load (i.e., joint dynamics) using wearable technology. This is of paramount importance, as ambulatory joint load assessment can improve health diagnostics and rehabilitation of injuries as well as musculoskeletal diseases. The overall aim of this thesis is to assess how, and to what extent, accelerometer sensor data and machine learning techniques can biomechanically quantify sports performance and the load on body structures during the execution of everyday and sport movements. This work proposes a conventional neuron network (CNN) for successful human activity recognition using Accelerometer. Various experiments are performed on a real-world wearable sensor dataset to verify the effectiveness of the deep learning algorithm. The results show that the proposed that CNN 1D model performs comparison with other CNN 2D model and achieves satisfactory activity recognition performance. Some open problems and ideas are also presented and should be investigated as future research.