Thanks Proga,
Luminar's patent for their tracking NN software was filed in mid-2020. It may be that the part of their lidar Volvo are using is the capability to focus laser pulses on objects of interest, but if, as you say, they are running it on Nvidia GPU, then the new graph neural network software could be included in an update after being verified for safety, as it is a critical safety function.
Running on Nvidia, it will use a lot more power and be significantly slower than Akida in classifying objects.
I think that the foveated lidar will probably allow autonomous driving in excess of 100 kph because it will enable the receiver to detect distant objects with greater certainty than vanilla lidar.
The further away an object is, the fewer lidar pulses strike it because the scan angle is quite broad (>100 degrees), so each pulse ray diverges from its adjacent pulses by 100/n degrees, where n is the number of pulses in a scan row. The laser pulse rays themselves do not get broader and spread out. This means that, the further away the object is, the larger the horizontal and vertical gaps between the pulses which hit the object.
Thus, by decreasing the angle between the pulse rays by increasing the rate at which the pulses directed to the objects of interest are generated while keeping the scan rate constant, more pulses strike the object and send reflections to the receiver.
There was a recent test quoted where Akida 1 proved to be 30 times better than Nvidia Jetson. I think tracking multiple objects using software on Nvidia will burn a lot of power.
grayscale.ai
.
