Neural-based Visual Odometry Trained in a Virtual Enviroment for a Mobile Robot Navigation

Visual odometry is a well-known technical problem, which is crucial for navigation of autonomous mobile robots in buildings or hard-to-reach areas where no precise external navigation is available. We propose an architecture of a neural network capable of solving the stated problem. Training of the network is performed based on data collected within a virtual environment. Additionally we suggest an algorithm for filtration of gathered data based on sensor fusion methods. The inputs of the whole system are two sequential videoframes and the subset of inertial sensor readings. The output of the proposed system is a vector of motion parameters containing information on robot's linear and angular velocities. The conducted experiments confirm adequacy and show possible application of the proposed network architecture and filtering algorithm for searching people lost in a forest.