Position independent activity recognition using shallow neural architecture and empirical modeling

Abstract

The goal of the SHL recognition challenge 2019 is to recognize transportation modalities in a sensor placement independent manner. In this paper, the performance of shallow neural networks is benchmarked by Team Orion in such a manner on the dataset provided in the challenge, using 156 handcrafted temporal and spectral features per sensor through the application of parallel processing and out-of-memory architecture. Using scaled conjugate gradient backpropagation (SCGB) algorithm, combining classes 7 and 8 and taking 5000 frames of bag-hips-torso data from validation set, classification accuracy of 87.2% was obtained on the validation dataset of the same labels for a shallow two-layer feed-forward network. 71% accuracy was obtained on the validation set of classes 7 and 8 via transfer of 2500 frames using another shallow neural network of similar architecture. Using empirically observed variable based transfer of 7088 frames from hand validation data to training dataset, 77.5% accuracy was obtained on hand validation data for classes 1 to 7/8, and 70% classification accuracy of classes 7 and 8 via transfer of 1809 frames from hand validation data. The results illustrate how carefully crafted features coupled with empirical transfer of labeled knowledge and combination of problematic classes can tune a neural classifier to work in a new feature space.