It’s been a long time to post (almost a year…).
There is a new selfie drone named Skydio R1 descendants of the former Google Project wing and MIT graduates and the video below shows what it can do.
It seems the drone does a pretty good job with 13 cameras (6 stereo pairs (i.e., one pair for each plane) global shutter camera +1 front filming high-res cam) with NVIDIA TX1 as the primary processing unit. Each stereo pair creates/estimates depths and represents them as voxel for local obstacle avoidance, motion planning, and object prediction (Skydio says 4 secs horizon). DNN is used for an object (mostly human) detection and tracking from the image taken from the front high-res camera (not sure!). More technical detail and explanations can be found from the interview with Adam Bry, who is the CEO of Skydio.
Here is more testing video. I don’t whether he got some support from Skydio but personally, I like his review that usually since he has frankly described gadget’s (DJI Mavic or GoPro) pros and cons in consumer’s perspectives.
If you want to know more about under-the-hood software packages, CMU RI seminar of Adam Bry would be a good starting point.