I’m pleased to announce that I’m now added support for several new yolo-based computer vision models to my home security project “StalkedByTheState” for the Jetson nano, Xavier NX and Xavier AGX including the current champion, YoloV7 (Chien-Yao Wang, Alexey Bochkovskiy, Hong-Yuan Mark Liao).
"YOLOv7 surpasses all known object detectors in both speed and accuracy..."
StalkedByTheState installs with just one command a full home security system with reverse proxy, web based security state machine, automatic renewed letsencrypt certificates, power-cut resilience running with a read-only OS mount (with memory FS overlay) and the most advanced yolo-based computer vision triggers and much more.
With one command sbts installs support for yolov7, scaled-yolov4, yolor, yolov4 and yolov4. The newest three models are pytorch/torchvision based models. The install script takes care of installing of the pre-requisites for the frameworks required for these models to use with the standard installed python3.6. After installation, pytorch/torchvision can be used without using conda python or virtualenvs which provides more flexibility for those that wish to install further non-ai focused python modules.
This software installs onto the Jetson nano, Jetson Xavier NX and Jetson AGX computers on top of jetpack 4.6.1.
The Jetson nano doesn’t have enough memory for the pytorch based models for this platform you can choose between yolov3 or yolov4 both are installed.
The Xavier NX (8GB machines) can run one of the pytorch-based versions, yolov7 is configured by default. But you can easily experiment with all of the other models which are also installed.
The Xavier AGX can support multiple models and SBTS can use different models for different parts of the image configurably as the models are accessed by a websocket interface added to the based software. By default yolov7 and yolov4 are started by default on this platform and others with more than 8GB of memory. For false positive prone areas it can be very effective to double check hits from yolov7 by confirming again with yolov4. This does reduce sensitivity for those areas but works very will for eliminating false positives for clustered parts of the image when using the pre-trained coco weights.
Next month I will start work on adding support for Jetpack 5.0* and the Jetson AGX Orin developer kit and will endeavour to add a system architecture diagram and more documentation to the github repo in the meantime.
The github repository is located here:
The software is designed to install on either an nvme SSD or USB connected SSD it won’t install to the SD card.
I hope you have fun playing with this. There is a reference to a YouTube channel showing installation and configuration in the github repo and I starting to add documentation to the github repository, starting with the architecture diagram. I realize the videos are too long.