Biomechanical Markerless Motion Classification Based On Stick Model Development For Shop Floor Operator

Motion classification system marks a new era of industrial technology to monitor task performance and validate the quality of manual processes using automation. However, the current study trend pointed towards the marker-based motion capture system that demanded the expensive and extensive equipment setup. The markerless motion classification model is still underdeveloped in the manufacturing industry. Therefore, this research is purposed to develop a markerless motion classification model of shopfloor operators using stick model augmentation on the motion video and identify the best data mining strategy for the industrial motion classification. Eight participants within 23 to 24 years old participated in an experiment to perform four distinct motion sequences: moving box, moving pail, sweeping and mopping the floor, recorded in separate videos. All videos were augmented with a stick model made up of keypoints and lines using the programming model. The programming model incorporated the COCO dataset and OpenCV module to estimate the coordinates and body joints for a stick model overlay. The data extracted from the stick model featured the initial velocity, cumulative velocity and acceleration for each body joint. Motion data mining process included the data normalization, random subsampling method and data classification to discover the best information for separating motion classes. The motion vector data extracted were normalized with three different techniques: the decimal scaling normalization, min-max normalization, and Z-score normalization, to create three datasets for further data mining. All the datasets were experimented with eight classifiers to determine the best machine learning classifier and normalization technique to classify the model data. The eight tested classifiers were ZeroR, OneR, J48, random forest, random tree, Naïve Bayes, K-nearest neighbours (K = 5) and multilayer perceptron. The result showed that the random forest classifier scored the best performance with the highest recorded data classification accuracy in its min-max normalized dataset, 81.75% for the dataset before random subsampling and 92.37% for the resampled dataset. The min-max normalization gives only a slight advantage over the other normalization techniques using the same dataset. However, the random subsampling method dramatically improves the classification accuracy by eliminating the noise data and replacing them with replicated instances to balance the class. The best normalization method and data mining classifier were inserted into the motion classification model to complete the development process