Abstract: The disclosed technology provides solutions for confirming a sensor cleaning event of a sensor mounted on an autonomous vehicle. A method of the disclosed technology can include steps for receiving a first plurality of image frames from an optical sensor; segmenting each of the first plurality of image frames into one or more interest regions; computing image statistics for each of the one or more interest regions; generating an image mask based on the image statistics for each of the one or more interest regions; receiving a second plurality of image frames from the optical sensor; applying the image mask to the second plurality of image frames; computing image statistics for each of the second plurality of image frames; and determining a status of a sensor cleaning event based on the image statistics for each of the second plurality of image frames. Systems and machine-readable media are also provided.

Abstract: The present technology determines noise in an image of a representation of an environment around an autonomous vehicle (AV). A sensor system receives a first image that represents an environment around the AV including at least one object that may be obscured by environmental conditions. The first image is analyzed to determine if pixels in the first image represent noise from the environmental conditions in comparison to other pixels in the first image that represents a higher degree of image detail. A second image can be generated using the pixels in the first image that represent the higher degree of image detail to result in an output that is less affected by noise from environmental conditions and may result in better performance by downstream systems relying of date from the sensor system.

Abstract: Use of camera information for radio detection and ranging (RADAR) beamforming is disclosed. Camera information for a scene having a target object is received. Digital RADAR waveforms corresponding to a frame of reference including the at target object are received. Coordinates for the scene in the camera frame of reference are translated to the RADAR frame of reference. A radar cross section estimation is determined for the object based on the transformed coordinates. A kernel is selected based on the radar cross section estimation. RADAR signal processing is performed on the digital RADAR waveforms utilizing the selected kernel. A point cloud is populated based on results from the RADAR signal processing.