Enhanced Object Detection Using Random Forests and Image Augmentation Techniquess
Published:
Abstract
This project presents a classical object detection pipeline, implemented in C++ using OpenCV, tailored for three distinct classes. The methodology encompasses selective search for region proposals, Histogram of Oriented Gradients (HOG) for feature extraction, and random forests for window classification. The HOG descriptor is optimized through experimental tuning of parameters, including window size and block stride. An imbalanced dataset necessitates the random sampling technique to ensure class balance during the training of OpenCV’s DTrees decision trees. Non-Maximum Suppression (NMS) is applied to refine the final bounding boxes, with an overlap threshold of 0.3, evaluated through Intersection over Union (IoU) and Intersection over Max Area metrics. The pipeline demonstrates enhanced performance, with optimal results achieved using a random forest of 100 decision trees and a specifically parameterized HOG descriptor.