The human brain is the most complex, powerful and versatile learning machine ever known. Consequently, many scientists of various disciplines are fascinated by its structures and information processing methods. Due to the quality and quantity of the information extracted from the sense of sight, image is one of the main information channels used by humans. However, the massive amount of video footage generated nowadays makes it difficult to process those data fast enough manually. Thus, computer vision systems represent a fundamental tool in the extraction of information from digital images, as well as a major challenge for scientists and engineers.
This thesis' primary objective is automatic foreground object detection and classification through digital image analysis, using artificial neural network-based techniques, specifically designed and optimised to be deployed in low-cost hardware devices. This objective will be complemented by developing individuals' movement estimation methods by using unsupervised learning and artificial neural network-based models.
The cited objectives have been addressed through a research work illustrated in four publications supporting this thesis. The first one was published in the “ICAE” journal in 2018 and consists of a neural network-based movement detection system for Pan-Tilt-Zoom (PTZ) cameras deployed in a Raspberry Pi board. The second one was published in the “WCCI” conference in 2018 and consists of a deep learning-based automatic video surveillance system for PTZ cameras deployed in low-cost hardware. The third one was published in the “ICAE” journal in 2020 and consists of an anomalous foreground object detection and classification system for panoramic cameras, based on deep learning and supported by low-cost hardware.