Using machine learned features for robot ego-motion estimation through an event-camera

This thesis analyses the advantages offered by event-cameras in ego-motion estimation. Traditional cameras suffer from poor performance in low light conditions or high-speed motion. Event-cameras overcome these limitations by detecting and processing only the changes in the visual scene, offering a higher dynamic range and a lower power consumption. In particular, this thesis analyses a feature detection method based on machine learning that takes advantage of the peculiarities of this type of data, resulting in higher precision and longer feature tracks with respect to handcrafted methods. The inference pipeline is composed of a module repeated twice in sequence, formed by a Squeeze-and-Excite block and a ConvLSTM block with residual connection. It is followed by a final convolutional layer that provides the trajectories of the corners as a sequence of heatmaps. A novel training method is described and evaluated.