Motion compensation and refinement in recurrent neural networks

• Main Authors: KOSATSKY, IGOR, LABUSCH SVEN, MEUTER MIRKO
• Format: Patent
• Language: Chinese; English
• Published: 30.08.2022
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

The invention relates to motion compensation and refinement in recurrent neural networks. A computer-implemented method for tracking motion of information in an environment of a vehicle, the method comprising: acquiring sensor-based data about a spatial environment of a vehicle over a plurality of time steps, the sensor-based data defining information in respective spatially resolved cells of the spatial environment; for each time step of the plurality of time steps, inputting the sensor-based data into a recurrent neural network (RNN) having one or more internal memory states; transforming one or more internal states of the RNN for each time step of the plurality of time steps by using a motion map describing the velocity and/or direction of motion of the information of the respective spatially resolved cell based on the sensor data, respectively; and for each of a plurality of time steps, using the transformed internal state in the processing of the RNN to track the movement of information in the environmen