Determining ego motion

A method for training a neural network for determining ego motion of a vehicle, the method comprising:

• providing different frames of radar data of a training vehicle surrounding environment and at least one of corresponding localization or driving information of the training vehicle;
• deriving ego motion from the radar data by:
• acquiring different frames of radar data of the training vehicle surrounding environment, each frame being acquired at a different time;
• deriving from the radar data of each different frame, an environment map of the vehicle surrounding environment;
• and
• deriving the ego motion of the vehicle by:
• merging environment maps from at least two different frames into one accumulated map by providing a sequence of environment maps to one or more recurrent neural network layers, wherein the accumulated map is a digital image or file comprising a plurality of pixels and the one or more recurrent neural network layers comprise convolutional long-short term memory (LSTM) layers configured to encode spatial relationships and displacements between different time steps to generate the accumulated map;
• computing, from the accumulated map, a motion vector for each pixel of the accumulated map, wherein each pixel of the accumulated map is assigned two X values and two Y values defining the motion vector;
• and
• extracting, from the accumulated map, a mask map including a tensor mapping a weight for each pixel of the accumulated map, wherein the mask map is extracted by one or more neural network layers comprising one or more of a convolutional layer, a pooling layer, and a self-attention layer;
• deriving ground truth ego motion based on the at least one of localization or driving information;
• and
• determining optimal weights of a neural network by minimizing a loss function between the ego motion as an input and the ground truth ego motion as a target output.

A non-transitory computer-readable medium comprising instructions which, when executed by a computer, cause the computer to:

• acquire, from a vehicle radar emitter with a plurality of reception antennae, different frames of radar data of a vehicle surrounding environment, each frame being acquired at a different time, wherein the radar data includes a plurality of Doppler channels;
• derive, from the radar data of each different frame, an environment map of the vehicle surrounding environment by reducing the plurality of Doppler channels to one feature channel and resolving Doppler ambiguities by acquiring the different frames of radar data using different looktypes, wherein the different looktypes comprise different resolutions in range or resolutions in Doppler for different frames, and wherein the Doppler ambiguities are resolved by performing a temporal fusion of at least two detections of a same object across the different frames;
• and
• derive ego motion of the vehicle by:
• merging environment maps from at least two different frames into one accumulated map, wherein the accumulated map is a digital image or file comprising a plurality of pixels;
• computing, from the accumulated map, a motion vector for each pixel of the accumulated map, wherein each pixel of the accumulated map is assigned two X values and two Y values defining the motion vector;
• and
• extracting, from the accumulated map, a mask map including a tensor mapping a weight for each pixel of the accumulated map, wherein the mask map is extracted by one or more neural network layers comprising one or more of a convolutional layer, a pooling layer, and a self-attention layer.