Object removal using lidar-based classification

摘要

In scenarios involving the capturing of an environment, it may be desirable to remove temporary objects (e.g., vehicles depicted in captured images of a street) in furtherance of individual privacy and/or an unobstructed rendering of the environment. However, techniques involving the evaluation of visual images to identify and remove objects may be imprecise, e.g., failing to identify and remove some objects while incorrectly omitting portions of the images that do not depict such objects. However, such capturing scenarios often involve capturing a lidar point cloud, which may identify the presence and shapes of objects with higher precision. The lidar data may also enable a movement classification of respective objects differentiating moving and stationary objects, which may facilitate an accurate removal of the objects from the rendering of the environment (e.g., identifying the object in a first image may guide the identification of the object in sequentially adjacent images).

主权项

1. A computer-implemented method of generating a rendering of an environment including at least one object, the method performed on a device comprising a processor, the method comprising:
generating for the environment, a lidar point cloud comprising at least one lidar point captured over a period of time; mapping respective lidar points in the lidar point cloud to identify an object in the environment; selecting a movement classification of the object according to the at least one lidar point; based on the lidar points in the lidar point cloud, identifying the object in a first representation of the environment captured at a first time within the period of time; based on an identified position of the object at the first time and the movement classification, estimating a position of the object at a second time within the period of time; based on the estimated position, identifying the object in a second representation of the environment captured at the second time; and generating the rendering of the environment omitting at least a portion of the object, based at least on identification of the object in the second representation.