How the Dahua 5442T-ZE identifies people/cars

[d5775927]

1) Does anybody know what algorithm is used?
I did a bit research and using such algorithms require much more resources (for example, Yolo algorithm).
2) Assuming Yolo algorithm (or similar one) is used, how is it possible, that the camera is able to run the algorithm with its limited hardware (is there a GPU inside)?

 

[ajareg]

You can see GitHub - facebookresearch/Detectron: FAIR's research platform for object detection research, implementing popular algorithms like Mask R-CNN and RetinaNet. or similar projects like GitHub - AlexeyAB/darknet: YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet ) (needs GPU hardware)
Or hardware decisions like Coral Edge Tpu or Nvidia Jetson. It's allow to detect objects directly from the camera stream in realtime!

 

[wittaj]

You can see GitHub - facebookresearch/Detectron: FAIR's research platform for object detection research, implementing popular algorithms like Mask R-CNN and RetinaNet. or similar projects like GitHub - AlexeyAB/darknet: YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet ) (needs GPU hardware)
Or hardware decisions like Coral Edge Tpu or Nvidia Jetson. It's allow to detect objects directly from the camera stream in realtime!

This is for AI from a video stream.

The OP wants to know what internally is the Dahua AI being used - is it one of these, Deepstack, or something else based on Yolo or something proprietary to Dahua.

 

[ajareg]

This is for AI from a video stream.

The OP wants to know what internally is the Dahua AI being used - is it one of these, Deepstack, or something else based on Yolo or something proprietary to Dahua.

I think it's something proprietary and there is no sence to debug it, it's easie to take any TPU/GPU module and use something like GitHub - blakeblackshear/frigate: NVR with realtime local object detection for IP cameras

 

[d5775927]

Will answer myself after checking the CPU of the camera (HI3516DV300):

Yuanchuang Huawei HiSilicon Hi3516DV300 face recognition Ai-DV300 - 深圳市源创嵌入式科技有限公司

Yuanchuang Huawei HiSilicon Hi3516DV300 face recognition Ai-DV300 - 深圳市源创嵌入式科技有限公司

www.yc-arm.com

TL;DR: it has an hardware accelerated NPU which is about 30x stronger than the performance of a raspberry pi4 (Deep learning with Raspberry Pi and alternatives in 2022 - Q-engineering).
For the specific algorithm in use, guessing:

[1605.06409] R-FCN: Object Detection via Region-based Fully Convolutional Networks

based on:

GitHub - openhisilicon/HIVIEW: A multi-process software framework for hisilicon (海思) ipc/dvr/nvr/ebox

A multi-process software framework for hisilicon (海思) ipc/dvr/nvr/ebox - openhisilicon/HIVIEW

github.com