Structured Tracking for Safety

Emerging developments in the speed, size, and power requirements of processors, coupled with networking advances, enable new applications for networks of sensors, cameras, and robots. However, we live in a world filled with uncertainty and noise, which affects the sensors we use, the environments we model, and the objects we observe. In this dissertation, we define Structured Tracking, where we apply novel machine learning and inference techniques to leverage environmental and tracked-object structure. This approach improves accuracy and robustness while reducing computation. We focus on three application areas of societal benefit: safety, security, and privacy. We apply Belief Propagation (BP)[148] algorithms to sensor networks, and describe our modular framework for the more general Reweighted-BP formulation[203]. To improve safety, we track pallets in warehouses. We apply Particle Filtering[162] and model the cardiod-shaped response pattern of ultrasound between static beacons and mobile sensors to improve tracking accuracy by 11%. We also show using inter-distance sensor readings, we can improve accuracy by 3--4x over the recent SMCL+R formulation[46], while being more likely to converge. We use a generalization of Particle Filtering, Nonparametric-BP (NBP)[183], which can model multi-modal and ring-shaped distributions found in inter-distance tracking problems. We develop a novel tracking algorithm based on NBP to fuse dynamics and multi-hop inter-distance information that increases accuracy, reduces computation, and improves convergence. For security, we present a novel approach for intruder surveillance using a robotic camera, controlled by binary motion sensors and use Particle Filtering to model intruder dynamics and environment geometry. We also present a localization-free approach to robot navigation using a distributed set of beacons, which emit a sequence of signals to direct a robot to the goal, modeling the robot's dynamics uncertainty, with up to 93.4% success rate. We introduce an approach to privacy for visual surveillance, "Respectful Cameras," that uses probabilistic Adaptive Boosting[68] to learn an environment-specific 9-dimensional color model to track colored markers, which act as a proxy for each face. We integrate probabilistic Adaptive Boosting with Particle Filtering to improve robustness, and demonstrate a 2% false-negative rate.