Task:
- Setup board
to take a stream of the camera we need to create a file that includes the below code file will be provided.
vi postrouting.sh # input the code file
# :wq
we need to install iptables to use the code
sudo apt-get install iptables
After running the code we let the rtsp TCP and udp address send and allow to 554 port on other device that we intend to use.
Camera rtsp data
on the postrouting.sh code, we need to change four lines to change to another camera or in the beginning:
this is a range we set, which will be around 100 and 192.168.144; the 144 is related to camera stream RTSP
here, it is prerouting the stream rtsp to dport 554, so we need to change to 192.168.144.25:8554 (to the camera that we are using)
Here, we change that to 192.168.144.25, which is camera rtsp ip.
Full code in below:
chmod +x postrouting.sh
./postrouting.sh
-
connect camera
-
open in vlc
to open the stream on our vlc, we use the eth1 which is the static ip address of board as below
rtsp://192.168.0.38:554/main.264, the ip address is of ab300 board
necessary libraries of android for qcom snpe:
https://android.googlesource.com/platform/frameworks/native/+/android-4.2.2_r1/include/utils/
https://android.googlesource.com/platform/system/core/+/master/libcutils/include/cutils/properties.h
https://github.com/jungin500/snpe-docker - snpe docker
https://developer.qualcomm.com/forum/qdevnet-forums/computer-vision-fastcv/7556 - qcom build issues
https://colab.research.google.com/drive/1Y7hjXPiTgkO7tsgG2xGbWVlMTNmUTh6A#scrollTo=ArMQB8cbHf_Q&uniqifier=1 = onnx usage on colab
https://mpolinowski.github.io/docs/IoT-and-Machine-Learning/ML/2023-01-14-yolov7_to_tensorflow/2023-01-14/ - onnx to tflite
https://igor.technology/export-yolo-v7-to-tensorflow-lite/ = onnx-to-tflite
https://github.com/VikasOjha666/yolov7_to_tflite/blob/main/yoloV7_to_TFlite%20.ipynb = onnx-tflite
https://github.com/sithu31296/PyTorch-ONNX-TFLite = pt - onnx - tflite
- the default detect.tflite is working and files are being organized in like below yolo folder
all the below ouptuts are default object detection results in qcom real time streaming
and it has reached 30 + frame per second
in below yolo folder there is and .tflite weight, label file and .config file to be used in code
gst-launch-1.0 qtiqmmfsrc ! video/x-raw\(memory:GBM\), format=NV12, width=1280, height=720, framerate=30/1 ! queue ! qtimletflite config=/data/misc/camera/yolo/mle_tflite.config
postprocessing=detection ! queue ! qtioverlay ! waylandsink width=1920 height=1080 async=true sync=false enable-last-sample=false
- qtimletflite is working as above code
this one is the qcom rb5 camera input
and output
gst-launch-1.0 rtspsrc location=rtsp://192.168.0.38:554/live ! decodebin ! videoconvert ! qtimletflite config=/data/misc/camera/yolo/mle_tflite.config
postprocessing=detection ! queue ! qtioverlay ! waylandsink width=1920 height=1080 sync=true enable-last-sample=false
and this one is ab300 camera rtsp://192.168.0.38 stream input
and output
- bugs related links (onnx-tensorflow, )
https://github.com/onnx/onnx-tensorflow https://github.com/onnx/onnx-tensorflow/issues?q=is%3Aissue+is%3Aopen+Transpose onnx/onnx-tensorflow#310 (NCHW" but tensorflow data type is "NHWC") onnx/onnx-tensorflow#912 onnx（nchw）>> tf.pb(nchw) PINTO0309/PINTO_model_zoo#15 Is there an easy way to convert ONNX or PB from (NCHW) to (NHWC)? https://console.paperspace.com/github/gradient-ai/yolov5/blob/master/tutorial-paperspace.ipynb?machine=Free-GPU&file=%2F.pre-commit-config.yaml https://stackoverflow.com/questions/63880242/trying-to-convert-onnx-model-to-pb-model-using-onnx-tf-library
https://github.com/nnstreamer/nnstreamer-example/tree/main/bash_script/example_yolo