Trajectory reconstruction by means of an event-camera-based visual odometry method and machine learned features

Abstract. This paper presents a machine learned feature detector targeted to event-camera based visual odometry methods for unmanned aerial vehicles trajectory reconstruction. The proposed method uses machine-learned features to enhance the accuracy of the trajectory reconstruction. Traditional visual odometry methods suffer from poor performance in low light conditions and high-speed motion. The event-camera-based approach overcomes these limitations by detecting and processing only the changes in the visual scene. The machine-learned features are crafted to capture the unique characteristics of the event-camera data, enhancing the accuracy of the trajectory reconstruction. 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. In the experimental part, a sequence of images was collected using an event-camera in outdoor environments for training and test.