Enabling context aware data analysis for long-duration repetitive stooped work through human activity recognition in sheep shearing

There is evidence to suggest that changes in kinematics and neuromuscular control in activities that take place over long periods of time lead to increased injury risk. The collection of biometric data over long time periods could provide insight into these injuries. However, it is difficult to analyse long period biometric data for occupations as the analysis depends on the activity being performed, and it is not practical to manually label the amount of data required. A sufficiently accurate human activity recognition algorithm can provide a means to segment the activities and allow this analysis, but the classification must be robust to the inter-individual differences, as well as the intra-individual variations in movement over time that are the target of analysis. This work presents a person-independent human activity recognition algorithm for sheep shearing using a Hidden Markov Model with physical features that are identified to be relevant to spinal movement quality. The classifier achieved an F1 score of 96.47% in identifying the shearing task.