Classification of Human Activities Based on Automotive Radar Spectral Images Using Machine Learning Techniques: A Case Study

Target detection and classification is a very important topic in automotive environments, in order to effectively support the automatic driving systems. However, well known methodologies for target classification, as pedestrian activity recognition, often involve the use of algorithms with a high computational complexity. Starting from the micro-Doppler spectral images obtained from a low cost automotive radar we present three different methods for human activity recognition (HAR) and describe their pros and cons. We compare the application of a deep neural network with two other alternative methods for feature selection, i.e., principal component analysis (PCA) and parameter extraction from micro-Doppler maps, combined with machine learning techniques. We consider the case study of a dataset containing three types of activities related to people's walking velocity. From the results obtained, we show that an appropriate choice of the parameters to extract, which has a relatively small computational cost, allows to achieve an excellent precision of more than 94%, which is higher than that achieved by more complex methods.