Tracking Objects with several cameras using Deep Learning on Raspberry Pi

This thesis is interested in multi-agent distributed tracking, robust to agent failures and target occlusions. First, the architecture of a network of communicating agents is conceptualized. It is composed of two types of agents: agent-cameras and an agent-user. The former constitute the intelligent sensors, which gather and filter data on the environment, while the latter assemble the filtered data for the end-user. The presence of multiple agents allows to compensate for occlusions by offering a variety of views on the targets. Then, as the observations collected by the agents are noisy, the problem of noise reduction is examined, with an implementation of a distributed Kalman filter. Experiments show a slightly improved noise reduction compared to the local Kalman filtering. Finally, an extension is considered, where the agents are allowed to move, in an attempt to improve their view of the targets they track. Depending on the use-case, this can improve the quality of the tracking.