This repository will not be updated. The repository will be kept available in read-only mode.
Note: This repo has been updated to use Maximo Visual Inspection (formerly known as PowerAI Vision). Everything is the same except for the name itself.
Whether you are counting cars on a road or products on a conveyer belt, there are many use cases for computer vision with video. With video as input, automatic labeling can be used to create a better classifier with less manual effort. This code pattern shows you how to create and use a classifier to identify objects in motion and then track the objects and count them as they enter designated regions of interest.
In this code pattern, we will create a video car counter using Maximo Visual Inspection, OpenCV and a Jupyter Notebook. We'll use a little manual labeling and a lot of automatic labeling to train an object classifier to recognize cars on a highway. We'll load another car video into a Jupyter Notebook where we'll process the individual frames and annotate the video.
We'll use our deployed model for inference to detect cars on a sample of the frames at a regular interval. We'll use OpenCV to track the cars from frame to frame in between inference. In addition to counting the cars as they are detected, we'll also count them as they cross a "finish line" for each lane and show cars per second.
Credit goes to Michael Hollinger for his initial notebook counting objects with PowerAI Vision.
When the reader has completed this code pattern, they will understand how to:
This code pattern requires Maximo Visual Inspection (formerly known as PowerAI Vision).
Go here
to learn more about trial access (Scroll down to the Give it a try
section).
This code pattern has been updated with screenshots and instructions for PowerAI Vision 1.1.3, but the instructions are the same for IBM Maximo Visual Insights.
The code included in this code pattern runs in a Jupyter Notebook. The notebook itself does not require Maximo Visual Inspection or Power Systems (only access to the deployed API). To run the Jupyter Notebook locally, install it using Anaconda. The installation instructions are here.
Hint: If you need a shortcut, you can import the dataset from
data/examples/dataset_auto_labeled.zip
, train and deploy that dataset, and then run the notebook (but you'll get more out of this if you go through all the steps).
To create a new dataset for object detection training from a video, use the Maximo Visual Inspection UI and start with a small manually annotated dataset (we'll expand on it with automatic labeling later).
Download the video to use to train the dataset from here. Use the Download
button to create training_video.mp4
in your browser's Downloads folder.
Click on the Data Sets
tab.
Click on the Create new data set
card.
Provide a data set name.
Click the Create
button.
Click on the newly created data set card.
Click on Import files
and open your downloaded training_video.mp4
file.
Select the new uncategorized card and click the Label objects
button.
Click on Auto capture
, set the capture interval to 5 seconds, and click the Auto capture
button.
Tips! Click on
How to label
for more detailed labeling tips.
Create a new object label for the data set by clicking + Add new
by the Objects
list. For example, add a car
object label.
Label the objects in the frames by following these steps.
Box
from the bottom left.Train model
.Object detection
.Advanced
settings. You can train faster (with less accuracy) by reducing the max iterations.Train
.Models
tab.Deploy model
button.Deployed Models
tab to see the status.We use the first deployed model that you trained with manually annotated frames and use inference to automatically label more cars in your training video.
Label objects
button.Auto label
.Capture Interval (Seconds)
.Auto label
button.Frames are captured at the specified interval and labels are added by using the specified trained model. By default, the automatically added labels are light red. After processing, you can manually add labels to the frames that have been auto labeled and you can manipulate (move, resize) the labels that were automatically generated. If a frame with automatically generated labels is edited, all labels on the frame are converted to manual labels.
Repeat the above train and deploy process with the newest dataset which was enhanced with automatic labeling.
This dataset has many more frames and labeled objects. It will create a much more accurate model.
The code included in this code pattern runs in a Jupyter Notebook. After you configure the URL of your deployed model in the notebook, you can just run it, read it, and watch the results.
Start your Jupyter Notebooks. Starting in your powerai-counting-cars
cloned repo directory will help you find the notebook and the output as described below. Jupyter Notebooks will open in your browser.
cd powerai-counting-cars
jupyter notebook
Navigate to the notebooks
directory and open the notebook file named counting_cars.ipynb
by clicking on it.
Edit the cell below Required setup! to replace the URL with the API endpoint of your deployed model. Use the copy button to capture the whole URL.
Use the drop-down menu Cell > Run All
to run the notebook, or run the cells one at a time top-down using the play button.
As the cells run, watch the output for results or errors. A running cell will have a label like In [*]
. A completed cell will have a run sequence number instead of the asterisk.
The Test the API on a single frame cell will demonstrate that you have correctly deployed your inference API. It should output JSON that includes classified cars. A portion of the output would look something like this:
"classified": [
{
"confidence": 0.9997443556785583,
"ymax": 370,
"label": "car",
"xmax": 516,
"xmin": 365,
"ymin": 240
}
]
The Get object detection results for sampled frames cell runs inference on a sampling of the video frames. The output will show a progress counter like this:
The Inference, tracking, and annotation cell processes every frame and has a similar progress counter. You can also preview the annotated frames as they are created in the output
directory.
The Play the annotated frames in the notebook cell displays the annotated frames in a loop to demonstrate the new video after they are all created. The notebook animation is usually slower than the video.
You can create an MP4 video from the annotated frames if you have a working installation of ffmpeg. The command is commented out as the last cell of the notebook. You can run it from there, or use the script in tools/create_video.sh
. The script takes the output directory (with the annotated frames) as an argument like this:
cd powerai-counting-cars
./tools/create_video.sh notebooks/output
Note: There is also a tool to create a gif from the video. We used that to show the sample output below.
As the notebook cells run, check for errors and watch the progress indicators. After the video has been annotated, the frames will play (like a video) in the notebook. The notebook playback is usually slow. If you used ffmpeg
to create an annotated video, you can play it back at full speed.
Example annotated video: here
Example notebook with static output: here
Example compressed and converted to gif:
Stopped adding cars.
If you are using a trial environment, your model deployment may be limited to 1 hour. Simply deploy the model again and run the notebook over (or from where the errors started). Using cached results allows the notebook to continue where it left off.
This code pattern is licensed under the Apache License, Version 2. Separate third-party code objects invoked within this code pattern are licensed by their respective providers pursuant to their own separate licenses. Contributions are subject to the Developer Certificate of Origin, Version 1.1 and the Apache License, Version 2.