Investigation of tracking processes applicable to adjacent non-overlapping RGB-D sensors

The work presented in this thesis provides a framework for monitoring wide area indoor spaces built from multiple Microsoft Kinect sensors. A large field of coverage is achieved by placing the sensors in a non-overlapping configuration to reduce the interference between the projected structured patterns. A novel procedure is proposed for estimating the geometric calibration between sensors that enables a common representation for all data by providing many corresponding planes in the view volume of each sensor using a ‟paddle”. Within this framework, an investigation is conducted of diÔåÁerent depth-based spaces for people detection and tracking purposes. Kinect v.1 sensors bring a multitude of benefits to surveillance applications, mainly for occlusion reasoning. However, this sensor has important limitations in terms of resolution, noise and range. In particular, data becomes more scattered with distance along the optical axis of the camera resulting in non-homogeneous representations throughout the range. Furthermore, when considering the aggregated view, each camera produces a diÔåÁerent orientation of data. The polar coordinate space representation of the common ground plane is proposed that mitigates these limitations and eÔåÁectively aggregates the data from all sensors. The use of discriminative appearance models is a chief aspect in order to properly distinguish people from each other, especially where the density of people is high. A multi-part appearance model is presented in this work - the chromogram - which combines colour with the height dimension oÔåÁering high discriminative capabilities especially during occlusions periods. A critical stage for multi-target tracking systems is establishing the correct association between targets and measurements; also known as the data association problem. In this context, the data association stage is investigated by evaluating diÔåÁerent well known data association methodologies. An alternative tracking approach which does not require a data association process is also analysed - the Mean-Shift tracker. A modified version of the Mean-Shift tracker is proposed for tracking on the ground plane that integrates the use of chromograms that reduces distractions from the background and other targets. A new challenging dataset is proposed for the evaluation of multi-target tracking algorithms. The tracking methodologies proposed in this work are compared quantitatively in this framework.