A deep neural network for detecting obstacle instances using radar sensors in autonomous machine applications

摘要

Problem to be solved: to provide object detection for an autonomous machine using a deep neural network (DNN).DNN is trained to detect both mobile and stationary obstacles from 3D (3D) radar data in both arterial and urban scenarios.The radar detection is stored, corrected by the ego action, and is positively projected and supplied to the neural network.The DNN includes a common trunk having a feature extractor, and a number of heads predicting a different output such as a class trust head that predicts the reliability map and a detected instance return head.The output is decoded, filtered and / or clustered to form a boundary shape that identifies the position, size, and / or orientation of the detected object instance.An object instance detected is provided to an autonomous vehicle operation stack to enable safe planning and control of an autonomous vehicle.Diagram