Automated counting of animals using rotation invariant features and machine learning classifier
Project is structured as follows
Empty directory for storing the adaboost classifiers created by sklearn.
This code finds wildebeest in images. Loops through a directory of aerial images, first identifiers pixels that occupy regions above a threshold contrast level, tests whether each pixel is identified as the centre of a wildebeest, then groups contiguous blocks of positive identifications into a single wildebeest
Empty directory for storing images that have been counted and have identified wildebeest marked
This directory contains the classes to extract rotation invariant features. It is an implementation of the algorithm proposed in Liu, Kun, et al. "Rotation-invariant HOG descriptors using fourier analysis in polar and spherical coordinates." International Journal of Computer Vision 106.3 (2014): 342-364. Two versions of the algorithm exist, circularHOGExtractor calculates features assuming it is a centred image of a single object whereas cudaHOGExtractor produces an array of features with each pixel of the image considered as a potential centre for the object
Contains various utilities for creating a training data set (clicking on images and manually correcting a first-pass trained ML algorithm) and code for creating an adaboost classifier using the sklearn package