When STEREOLABS dropped by the NVIDIA booth, I was able to get a peek at their ZED Stereo Camera. At first glance, the camera looked to be ruggedly packaged and a good match for autonomous vehicles and drones.
In a very interesting Session Talk Stereo Vision For Autonomous Machines given by Edwin Azzam, the CTO of STEREOLABS, the pipeline for image rectification and depth mapping was outlined as implemented on a Jetson TK1.
I enjoyed the amount of information that was shared about the process.
The compute efficiency noted that it takes about 7 ms to upload the images, 3 ms to rectify them, and then 22 ms to calculate the depth map on the Jetson TK1 GPU. In the background, the CPU runs a calibration check program. Given that real time sensing and obstacle detection must occur in 50 ms, this leaves about 18 ms to be used for obstacle detection. This looks like a nice fit for the Jetson TK1. platform.
This looks to be a nice package for placing on robots and drones, and should be available soon.