Machine learning has recently made great progress in object classification, and especially object classification in images. In some cases, machine learning has even shown better ability than the human. There is a great potential to partially, or fully, automate sensor data analysis in surveillance systems. The automation could greatly facilitate the human operator’s work in finding critical information. After the success of object classification in images, the next step is to try to achieve similar progress in behaviour recognition using similar approaches. The paper includes a brief state-of-the-art for automatic behaviour recognition, and especially behaviours related to surveillance. The focus is on approaches based on deep learning. The paper presents an experiment with a specific deep learning algorithm, namely the long short-term memory (LSTM). For behaviours in surveillance applications, actual training data are rare. One approach for overcoming this is to use a deep learning algorithm that is pre-trained for a nearby application, and then transferred to the current application. Another is to expand the amount of training data using simulations. A difficulty with simulations is to create data that have similar characteristics as real sensor data, that is, include relevant noise and uncertainties. The paper briefly discusses the case of simulated data for the development of a deep learning method.
|