Video annotation framework for accelerometer placement in worker activity recognition studies

Abstract

Automated recognition of worker activities has the potential in aiding quick assessment of labour productivity on construction sites. A novel method called accelerometer based activity recognition has been investigated and preliminary results show that it has good potential for deployment in construction environment. The major decisive factor influencing the performance of the activity recognition system is the location of the accelerometer on the human body. The objective of this study is to determine a-priori, the appropriate accelerometer location using videos of construction activities. A framework was developed to track the movement of body segments through observation using Anvil, a generic annotation tool. Locations of accelerometer are selected after evaluating the information gain of each body segment towards activity classification with due consideration to subject comfort and integration possibilities. A study of masonry activity using the framework identified the placement locations as right lower arm, left lower arm and waist. An experimental setup was arranged for determining performance of the accelerometer based activity recognition system for a mason working in uninstructed environment with accelerometers attached at selected locations. Activity recognition performance of the locations was evaluated with ten runs of 10-fold cross validation using multilayer perceptron algorithm. The results showed that classifier performances for the three locations have the same order of ranking as predicted by the framework. The activity recognition performance for the selected locations gave accuracies above 80% and it can be concluded that the proposed framework can be used for placing accelerometers at appropriate locations for activity recognition.