Action Recognition from Low-Resolution Infrared Sensor for Indoor use: A Comparative Study between Deep Learning and Classical Approaches

In recent years, automatic action recognition has attracted a lot of attention in the field of computer vision. In this paper, action recognition in an indoor environment using low-resolution sensor is considered. In this environment, several studies have been carried out in visible imagery. Despite impressive performances, the main pitfall faced by the techniques developed is a marked restraint of the users to be filmed. To cope with the problems of personal identity revealed during the surveillance, we opted within the CoCAPS project for using a low resolution (64 x 64 pixels) infrared sensor with ceiling position which guarantees the respect of the intimacy of the person. Defined in collaboration with the industrial partners involved in the CoCAPS project, the scenarios targeted concern office situations and 7 classes of action have been selected, namely: no action, restlessness, sitting down, standing up, turning sitting on a seat, slow walking, fast walking. Classical approaches such as those based on the computation of shape descriptors (such as geometric and Color Histogram of Oriented Phases (CHOP)) extracted from Motion History Image (MHI) are investigated to represent action video sequences. Within this first group of classical approaches, the performance of a proposed model based on statistical attributes constructed from the tracking of centers of gravity of segmented forms is also presented. The comparative study is then completed by considering other models extracted from deep learning literature (convolutional neural networks (3D-CNN), Long Short Term Memory (LSTM)). The results obtained from the comparative study show that the proposed model is very competitive and provides promising results (83% of f-score).