• background subtraction
• local features
• object matching
• visual object similarity

In this work, a system for incremental real-time visual object detection and autonomous recognition is presented. The system is designed for indoor smart cameras and identifies objects appearing on the scene by detecting video changes in the video stream. Object detection is based on a novel interest point-based background subtraction method, which results in a more robust and informative
background model with respect to typically color-based approaches. Objects are incrementally learnt by collecting observations in real-time. A similarity function between objects observations relying on local feature matching and geometric consistency checking is defined. The key idea of the system is to relate past and present object observations: clusters of similar observations
are maintained exploiting transitivity of similarity between observations and are used to recognize a new observation of an already seen object. Since the system incrementally builds it up from observations during time, no training set for recognition is needed. Experiments have been performed on publicly available datasets to evaluate the detection task and the ability of the system to build good clusters of observations. The system has also been tested on the Raspberry Pi platform equipped with the Pi Camera module.