One of the typical characteristics of children with developmental disabilities is stereotypic behavior, which is the repetition of the same movement. Stereotypic behavior could be a barrier to the social integration of children with developmental disabilities. One efficient key to solving this problem is applying Human Activity Recognition (HAR). In this study, we introduce a method for optimizing a wearable system for HAR. The body part to be analyzed is the finger, the most frequently moved joint, so the wearable device used is a glove with a 2-axis flex sensor and a processor. Since the HAR of this application must continuously be operated, a lightweight system is required. Furthermore, due to the characteristics of a wearable device that operates with a battery, it is essential to implement a long-running system by considering energy consumption and performance. To achieve this, we introduce a lightweight method in all stages from data collection to classifier. We propose a method to improve performance while minimizing the model size by designing Multi-Layer Perceptron (MLP)-based sequential classifier. A sequential classifier is suitable for resolving performance degradation caused by the similarity between gestures. First, classes corresponding to similar gestures are designated as an uncertain group. Then, if the output of the first classifier belongs to the uncertain group, the second classifier with a smaller size classifies it again. Due to the proposed method, higher performance could be achieved than when using a single classifier. As a result of the experiment, it is possible to achieve similar performance with a model with 72% fewer parameters than the optimization design achieved in the previous study.

Developmental disabilities; Lightweight Neural Network; Flex sensor; Hand motion Recognition; Human Activity Recognition