Haar Cascade for Object Detection

This work utilizes a Haar Cascade Classifier for the identification of wood logs. Positioned at the harvesting station, a camera is affixed to the harvesting machine, capturing images of the wood log cross-section whenever it aligns with the lens. The camera activation is synchronized with the movement of logs on the cutting belt. However, external factors occasionally lead to false triggers, resulting in the capture of images without wood logs. The Haar Cascade model is employed to discern the presence of wood logs in the images, effectively filtering out those without wood logs.

In this project, we implement a Haar Cascade using OpenCV. This is a machine-learning object detection method used to identify objects in images or video. It is based on the Haar Wavelet technique, which is a mathematical concept used in signal processing. The Haar Cascade is particularly popular for detecting faces in images.

Overview

1. To use a Haar Cascade Classifier for a specific object detection task, the classifier needs to be trained. During training, the classifier is provided with positive and negative samples of the object. Positive samples contain the object of interest, while negative samples do not.
2. The Haar Cascade method works by evaluating rectangular features at different scales in an image. These features are like filters that respond to specific patterns of intensity changes in the image.
3. To efficiently compute these features, the integral image concept is used. The integral image allows for the rapid calculation of the sum of pixel values in a rectangular region.
4. Adaboost Training: During training, a machine learning algorithm called Adaboost is used to select the most relevant features and assign weights to them. Weaker classifiers are combined into a strong classifier that can effectively discriminate between positive and negative samples.
5. During the detection phase, a sliding window is used to scan the image at different scales and positions. At each step, the classifier evaluates the features within the window to determine if the object is present.

Getting started

• Install OpenCV for annotation and training in Windows. Download the .exe file of the latest version of OpenCV here: https://sourceforge.net/projects/opencvlibrary/files/opencv-win/3.4.3/

• After downloading, extract the folder called 'opencv' and copy this path: C:\Users\username\Downloads\opencv\build\x64\vc15\bin

• In the above extracted folder, we are going to use opencv_annotation.exe, opencv_createsamples.exe, opencv_traincascade.exe and opencv_visualisation.exe. Please refer to the command.txt file for using the opencv commands.

• Put the positive images containing the object in the positive folder and the negative images without the objects in the negative folder respectively.

• cascadeutils.py: This script is used to create a list of negative files and store it in the neg.txt and the pos.txt file is generated from the OpenCV command opencv_annotation.exe:

  $ C:\opencv\build\x64\vc15\bin\opencv_annotation.exe --annotations=pos.txt --images=positive -m=1 -r=3 

• Next we need to create samples for the positive images with the following command:

  $ C:\opencv\build\x64\vc15\bin\opencv_createsamples.exe -info pos.txt -h 48 -w 48 -num 1000 -vec pos.vec 

  It creates samples and stores them in the pos.vec file which is a binary image file with width '-w' and height '-h' of the detection window size, and the number of vectors for '-num'.

• Next, train the classifier as shown below:

  $ C:\opencv\build\x64\vc15\bin\opencv_traincascade.exe -data cascade/ -vec pos.vec -bg neg.txt -w 48 -h 48 -numPos 8
    0 -numNeg 200 -numStages 10 

• Finally, the model is tested with main.py.

Demo Result

This is a test output after training the model with the company dataset from the forest and the test image is obtained from Google.

Dataset Notice

Please note that the dataset used in this project is proprietary and owned by a company client of Fraunhofer IPM. Unfortunately, we are unable to provide public access to the dataset due to legal and confidentiality agreements. The focus of this repository is to showcase the code, methodology, and results derived from the dataset.

References

Published Paper: https://www.cs.cmu.edu/~efros/courses/LBMV07/Papers/viola-cvpr-01.pdf

Link to the OpenCV tutorial: https://www.cs.cmu.edu/~efros/courses/LBMV07/Papers/viola-cvpr-01.pdf