Fully End-to-End Visual Odometry of a Minidrone

This paper deals with the integration of visual odometry of drones with neural networks. CNN is used in this paper to extract the feature encoding of a particular frame in time and then use RNN to output the position of the drone that is time-dependent. This is a complete end-to-end deep RCNN approach and no pre-processing of the image is required in this, only resizing of the image has to be done as per the model input. The model is trained on both simulation and real-world data, although the amount of data from the simulation is more. In this paper, the variations in the height of the drone are considered to a very less extent. Various models were compared based on metrics in order to pick the model with the best predictions and implemented on the mini drone. The best performance is obtained in the variable sequence length model using transfer learning.