Sparse Feature Extraction for Activity Detection Using Low-Resolution IR Streams

In this paper, we propose an ultra-low-resolution infrared (IR) images based activity recognition method which is suitable for monitoring in elderly care-house and modern smart home. The focus is on the analysis of sequences of IR frames, including single subject doing daily activities. The pixels are considered as independent variables because of the lacking of spatial dependencies between pixels in the ultra-low resolution image. Therefore, our analysis is based on the temporal variation of the pixels in vectorised sequences of several IR frames, which results in a high dimensional feature space and an "n<; <; p" problem. Two different sparse analysis strategies are used and compared: Sparse Discriminant Analysis (SDA) and Sparse Principal Component Analysis (SPCA). The extracted sparse features are tested with four widely used classifiers: Support Vector Machines (SVM), Random Forests (RF), K-Nearest Neighbours (KNN) and Logistic Regression (LR). To prove the availability of the sparse features, we also compare the classification results of the noisy data based sparse features and non-sparse based features respectively. The comparison shows the superiority of sparse methods in terms of noise tolerance and accuracy.