Evolutionary Approximation of Gradient Orientation Module in HOG-based Human Detection System

The histogram of oriented gradients (HOG) feature extraction is a computer vision method widely used in embedded systems for detection of objects such as pedestrians. We used Cartesian genetic programming (CGP) to exploit the error resilience in the HOG algorithm. We evolved new approximate implementations of the arctan function, which is typically employed to compute the gradient orientations. When the best evolved approximations are integrated into the SW implementation of the HOG algorithm, not only the execution time, but also the classification accuracy was improved in comparison with the accurate implementation and the state-of-the-art approximate implementations.