Comparison between digital image processing and learning techniques on a door detection task

Abstract

Door detection is a core task for any indoor autonomous robot at any level, such as assistive and domestic robots. However, the application of machine learning techniques can face hard constraints for embedded devices or autonomous platforms such as processing, memory, and energy. This paper aims to analyze and discuss the viability of two different approaches to the problem, a straight Digital Image Processing technique, and a Machine Learning based technique. Both strategies got compared under an Nvidia Jetson Nano. Therefore, after analyzing the resultant data, the Machine Learning based technique has shown the best overall mAP when compared to the Digital Image Processing technique (0.45 and 0.21, respectively). Also, the Machine Learning approach has proven to be more stable time-wise, which is another positive feature.