The objective is to achieve real time object detection at edge(vehicle counting/RLVD/ANPR/smart parking). We intend to connect 4 to 6 ip cameras(cat 6 cable preferably) on single Hardware. This requires at least 2 to 5 teraflops computation power to process atleast 10fps videos. The embedded board is to be deployed in outdoor installation and should be rugged enough to survive Indian conditions for smart city applications. It can be combination of either SBC with pci express slot to accommodate nvidia gtx970 or higher or can be combinatorial board like jetson tx2 developer kit.
We tried raspberry pi with movidius NCS but it isnt satisfactory.
Jetson TX2 is very suitable for such deployments. You can make a ruggedized sol with Jetson.
Also look at DeepStream SDK - [url]https://developer.nvidia.com/smart-cities[/url] to help you build your sol.
Something you may want to keep in mind with regard to Jetsons versus desktop PCIe video cards…the Jetsons already have a GPU capable of using a pre-trained model, but are not so good as a training device.
The GPU on the Jetson is integrated directly to the memory controller and is not PCIe. Most (all?) of the video drivers you find for the PCIe GeForce video products are not for arm64/aarch64/ARMv8-a, and since the existing TX2 driver is not for PCIe, this implies using a GeForce board such as the gtx970 on the Jetson may not be possible. Even if you could do this it would destroy the ability to use the Jetson without a much bigger power supply. You will be better off training using something like the gtx970 on a desktop (PCIe with lots of power required), and then deploying the model to a TX2 (very little power required, wired straight to memory controller) without any kind of external GPU.