EdgeTpu Detection Camera

This is an example of using AI to AI AIs so you can AI.

Requirements:

• Coding hoodies
• A linux machine, I tested this build on x86_64, armv7l, and aarch64 architecture.
• A camera to take inputs and a monitor to show frames.
• [Optional] An edgetpu device for speedup inference.

Dependencies Installation:

$ sudo apt install libgtk2.0-dev
$ sudo apt search libgtk2.0-dev
[Optional] install cmake-3.17 if needed
$ bash scripts/install_cmake.sh

Build

$ mkdir build && cd build
$ cmake ..
$ make

Notes:

• Since this builds libopencv*, libtensorflow-lite.a, libabsl*, and libglog, before building the project. It requires about ~7GB of storage, so you'll need an sdcard for the dev board. It will takes about two hours on the devboard and the rpi4 on the first build.
• If you encouter OOM killer, some swapspace (necesssary for devboard and rpi4):

$ bash scripts/make2gbswap.sh

Example Run

• CPU:

$ bin/{CPU}/edge --model_path test_data/mobilenet_ssd_v2_coco_quant_postprocess.tflite --label_path test_data/coco_labels.txt --height 480 --width 640

• EdgeTpu:

$ export LD_LIBRARY_PATH=libedgetpu/direct/{CPU}
$ bin/{CPU}/edge --model_path test_data/mobilenet_ssd_v2_coco_quant_postprocess_edgetpu.tflite --label_path test_data/coco_labels.txt --edgetpu --height 480 --width 640

Here are the list of options:

An example of using opencv with tflite/edgetpu in c++.
Usage:
  edge [OPTION...]

      --model_path arg    Path to .tflite model_file
      --label_path arg    Path to label file.
      --video_source arg  Video source. (default: 0)
      --threshold arg     Minimum confidence threshold. (default: .5)
      --verbose           To run in verbose mode.
      --edgetpu           To run with EdgeTPU.
      --height arg        Camera image height. (default: 480)
      --width arg         Camera image width. (default: 640)
      --help              Print Usage

Python

In case you didn't know, you can write this entire project in python with 71 lines of python... Notes: Requires installation of tflite_runtime-2.1.0.post1

• Run with CPU:

$ python3 scripts/camera_detector.py --model test_data/mobilenet_ssd_v2_coco_quant_postprocess.tflite --labels test_data/coco_labels.txt

• Run with TPU:

$ export LD_LIBRARY_PATH=libedgetpu/direct/{CPU}
$ python3 scripts/camera_detector.py --model test_data/mobilenet_ssd_v2_coco_quant_postprocess_edgetpu.tflite --labels test_data/coco_labels.txt --edgetpu True

Credits

Huge thanks to the following repos:

• google-coral/edgetpu taught me how to run inference in c++
• akioolin/edgetpu_demo taught me opencv
• powerluv/edgetpu taught me how to make the single CMakeLists.txt file

Because a wise man once told me: "Best documentation are code examples".