Classification of Human Indoor Activities with Resource Constrained Network Architectures on Audio Data

Abstract

Many functions and processes in Smart Homes can be improved by incorporating knowledge about the activities of the inhabitants. The inhabitant activities can be recognized by classifying typical sounds related to the activities using audio classification methods. In this paper, we present and approach for the classification of activity-based audio data, taking into account that in many cases the data processing has to be deployed on a resource-constrained device. An indoor dataset is created by recording 11 distinct human actions that are typically perfomed in a living room and kitchen setup. A Convolutional Neural Network (CNN) is used for classification that achieves an accuracy of up to 98%. In addition, network optimization methods like depthwise separable convolutions, pruning, channel scaling and quantization is used to reduce the memory footprint and computational requirements to obtain an optimized network with 93% fewer Million Floating point Operations (MFLOPs) and less than 2% drop in accuracy is achieved for use in resourceconstrained devices.