jc-cr / example-object-tracker

Edge TPU Object Tracker Example

Fork and modifications of Coral examples. Modified to do demo of delective object tracking and centroid depth extraction.

Installation

1. First, be sure you have completed the setup instructions for your Coral device. If it's been a while, repeat to be sure you have the latest software.

   Importantly, you should have the latest TensorFlow Lite runtime installed (as per the Python quickstart).

2. Clone this Git repo onto your computer:

   git clone https://github.com/jc-cr/example-object-tracker.git
   
   cd example-object-tracker/
   

3. Download the models:

   sh download_models.sh
   

   These models will be downloaded to a new folder models.

Run the detections

Importantly, you should have the latest TensorFlow Lite runtime installed (as per the Python quickstart). You can check which version is installed using the pip3 show tflite_runtime command.

1. CD into the gstreamer folder

   cd gstreamer
   

2. Install the GStreamer libraries and Trackers:

   bash install_requirements.sh
   

3. Run the detection model with Sort tracker

   python3 detect.py --tracker sort --target person --threshold 0.25 --videosrc /dev/video4
   

In the above command we use /dev/video41 to access the RGB stream from Intel 435i. If usign other depth camera, you could find available video sources using the command v4l2-ctl --list-devices --verbose`

Contents

• gstreamer: Python examples using gstreamer to obtain camera stream. These examples work on Linux using a webcam, Raspberry Pi with the Raspicam, and on the Coral DevBoard using the Coral camera. For the former two, you will also need a Coral USB Accelerator to run the models.

  This demo provides the support of an Object tracker. After following the setup instructions in README file for the subfolder gstreamer, you can run the tracker demo:

Models

For the demos in this repository you can change the model and the labels file by using the flags flags --model and --labels. Be sure to use the models labeled _edgetpu, as those are compiled for the accelerator - otherwise the model will run on the CPU and be much slower.

For detection you need to select one of the SSD detection models and its corresponding labels file:

mobilenet_ssd_v2_coco_quant_postprocess_edgetpu.tflite, coco_labels.txt