| Summary: | More complex image understanding algorithms are increasingly practical in a host of emerging applications. Object tracking has value in surveillance and data farming; and object recognition has applications in surveillance, data management, and industrial automation. In this work we introduce an object recognition application in automated nutritional intake analysis and a tracking application intended for surveillance in low quality videos. Automated food recognition is useful for personal health applications as well as nutritional studies used to improve public health or inform lawmakers. We introduce a complete, end-to-end system for automated food intake measurement. Images taken by a digital camera are analyzed, plates and food are located, food type is determined by neural network, distance and angle of food is determined and 3D volume estimated, the results are cross referenced with a nutritional database, and before and after meal photos are compared to determine nutritional intake. We compare against contemporary systems and provide detailed experimental results of our system's performance. Our tracking systems consider the problem of car and human tracking on potentially very low quality surveillance videos, from fixed camera or high flying \acrfull{uav}. Our agile framework switches among different simple trackers to find the most applicable tracker based on the object and video properties. Our MAPTrack is an evolution of the agile tracker that uses soft switching to optimize between multiple pertinent trackers, and tracks objects based on motion, appearance, and positional data. In both cases we provide comparisons against trackers intended for similar applications i.e., trackers that stress robustness in bad conditions, with competitive results.
|
|---|
