Detecção de Obstáculos com Câmeras Monoculares por meio de Mapas de Disparidades U/V e Redes Neurais Artificiais

Abstract

Environment perception is one of the most complex tasks to be performed au- tonomously. Besides depending on expensive sensors, many applications require high compu- tational power, limiting the applicability of such solutions. This paper presents a solution for environment perception with monocular cameras, focusing on low processing costs for embedded systems or modern smartphone applications. The solution uses classical image analysis tech- niques (such as disparity maps and their U/V variants) with modern techniques (deep artificial neural networks) to detect and estimate the distance of objects in space. Experimental results show good accuracy in distance estimation, significant improvements in object detection, and a runtime close to that of algorithms using only classical methods.