I did use Rekognition earlier and it was working fine & fast. However, back then I wrote a python script that took the images and pushed them to AWS from my video system
Since then I have skipped that. Benefit is also that you do not provide images to AWS. Instead I do local analysis using YOLO V3 full version. It is to my opinion very reliable but it depends I guess on what objects you target to detect. And it is good enough if you are checking individual images and have a reasonable powerful computer (I run it on an older good-enough laptop). But if you are looking for real time video streaming, I think you have to look for computers utilizing GPU's like Jetson Nano and it's inference engine or other usb stick based devices
If you just have the need to analyze images on certain (intrusion etc etc) events, you should certainly read this topic Object Detection using node-red-contrib-tfjs-coco-ssd
