Parameter selection for efficient HOG-based car detection

In this paper we explore how some parameters, less investigated in the literature, can be selected in a Histogram of oriented gradients (HOG)-based car detection system. The main goal is to find ways for reducing computation complexity while maintaining a good classification performance, in order to make possible real-time, embedded implementations. We analyze the effect of cell size, of binary representation of images and also of a new representation of the descriptor values, on fixed point, on only few bits. The fact that the simplified descriptor representation has overall the same performance as the floating-point one, opens the way to more aggressive fixed-point implementation of the HOG calculation.