It is important that autonomous vehicles quickly detect other cars or pedestrians that they share the road with. Researchers at Carnegie Mellon University have shown that they can significantly improve ‘detection accuracy’ by helping the vehicle recognize objects they don’t see.
Objects within your field of vision naturally prevent you from seeing things behind or further, but Peiyun Hu, a PhD student at the CMU Robotics Institute, explained why this is not true for autonomous vehicles.
Unlike a person, the presence of objects around autonomous vehicles allows him to work more accurately. The tools define objects as ‘cloud’ using data from LIDAR sensors, then try to match these clouds against objects in the 3D data library, but the part that Hu wants to draw attention to is; The 3D data from the sensors may not be really 3D. According to the doctoral student’s description, the sensor of the vehicle may not be able to see facades outside the field of view of an object, and existing algorithms cannot reason in such situations.
‘Perception systems must learn their own unknown’
According to Hu’s study, the detection system of an autonomous car operates based on ‘visibility’. In fact, this working principle was also used by companies to create digital maps. Professor Deva Ramanan from the CMU Argo Artificial Intelligence Center said, “The principle of mapping is based on free space and full space, but it is not always possible to handle obstacles that move depending on the speed of traffic.”
Hu and his colleagues used the techniques used in map making to help improve the stage of identifying objects in the system’s working principle of visibility. The team will present their project at the Computer Vision and Pattern Recognition (CVPR) conference, which will take place from 13-19 June.
The new CMU method developed performed better than the previous best performing technique when tested against a standard assessment. Method; improved his perception skill by 10.7% for cars, 5.3% for pedestrians, 7.4% for trucks, 18.4% for buses and 16.7% for trailers.
The reason for the lack of visibility of previous systems may be the concern about computation time, but Hu and his team say that this is no longer a problem because its methods take only 24 milliseconds to work. Each scan of LIDAR took 100 milliseconds.