Automatic Classification of Working Activities for Risk Assessment in Large-Scale Retail Distribution by Using Wearable Sensors: A Preliminary Analysis

Providing reliable information on human activities and behaviors is an extremely important goal in various application areas such as healthcare, entertainment, and security. Within the working environment, a correct identification of the actual performed tasks can provide an effective support in the assessment of the risk associated to the execution of the task itself, and thus preventing the development of work-related musculoskeletal diseases. In this perspective, wearable-based Human Activity Recognition systems have been representing a prominent application. This study aimed to compare three different classification approaches appointed from supervised learning techniques, namely k-Nearest Neighbors, Support Vector Machine and Decision Tree. Motion data, related to several working activities realized in the large-scale retail distribution, were collected by using a full-body system based on 17 Inertial Measurement Units (MVN Analyze, XSens). Reliable features in both time- and frequency-domain were first extracted from raw 3D accelerations and angular rates data, and further processed by Principal Component Analysis, with 95% threshold. The classification models were validated via 10-fold cross-validation on a defined training dataset. k-Nearest Neighbors classifier, which provide the best results on the training session, was eventually tested for generalization on additional data acquired on few specific tasks. As a result, considering 5 main macro activities, k-Nearest Neighbors provided a classification accuracy of 80.1% and a computational time of 1865.5 s. To test the whole assessment process, the activities labelled by the classification model as handling of low loads at high frequency were automatically evaluated for risk exposure via OCRA Checklist method.