An Easy-Parallelizable Evolutionary Optimization Algorithm via Probabilistic Bit-Wise Operation for Template Matching

Abstract

Pixel-based dense template matching still plays an important role in object localization tasks especially in the case of feature-less sequences. Instead of exhaustive search methods, optimization-based search strategy can converge to the global optimum faster by avoiding unnecessary tests. However, due to the presence of local optimums, conventional methods usually fail to converge in practice. In this paper, in order to prevent the optimization procedure from falling into a local optimal solution, we introduce a novel evolutionary operation called probabilistic bit-wise operation (PBO) into the framework of genetic algorithm (GA). Specifically, utilizing a natural phenomenon that the change of a higher-bit in an individual affect more than the lower-bit, the diversity of a population-based evolution algorithm can be well controlled by flipping each bit with different probabilities being assigned. Also, unlike crossover in GA, PBO only requires a single individual as the input, thus the parallelization can be easily implemented without considering the dependence between individuals. In the experiment, we compare against several classic optimization methods such as particle swarm optimization, GA, particle filter to show the superiority of PBO with a specific simulated benchmark. We also apply our method to a real sequence to show the practicality.