Human detection on Amlogic S905x devices using MS COCO 2017 datasets. Datasets were filtered using fiftyone library so that it only contains 'person' class. Model used is SSD MobileNetV2 and YOLOv4-tiny. The detection result can be viewed to website on this repository.
This guide assumes you have installed armbian to your machine. If you haven't, refer to this repository
Follow this step if you want to train the model from scratch. Otherwise, use the trained model provided on ssd and yolo folder
Use fiftyone to download datasets of certain class. Refer to fiftyone_coco.ipynb and adjust the path to your own machine.
Model are trained on google colab. Use techzizou's tutorial for MobileNetV2 and [YOLOv4-tiny](https://techzizou.com/train-a-custom-yolov4-tiny-object-detector-using-google(-colab-tutorial-for-beginners/ ). The train configs used to train the model on this repository are this file for MobileNet V2 and [this file] for YOLOv4-tiny.
Clone or download this repo to your Amlogic S905x device To run human detection inference on s905x devices:
main.py --model yolo --api http://localhost:5000/api/footage --url rtsp://KCKS:majuteru5@10.15.40.48:554/Streaming/Channels/1101 --cctv 2
. You can adjust the --model
to yolo
or mobilenet
. --api
used for API endpoint /api/footage
that this website is hosted on (adjust to your IP accordingly). url
used for RTSP URL. --cctv
used for CCTV number registered on database. http://localhost:5000/video_feed/[cctv_id]
Improving the model performance by using this repository