Vision-Based Target Objects Recognition and Segmentation for Unmanned Systems Task Allocation

This paper investigates the potential of deep learning methods to detect and segment objects from vision sensors mounted on autonomous robots to support task allocation in unmanned systems. An object instance segmentation framework, Mask R-CNN, is experimentally evaluated and compared with previous architecture, Faster R-CNN. The former model adds an object mask prediction branch in parallel with the existing branches for target objects location and class recognition, which represents a significant benefit for autonomous robots navigation. A comparison of performance between the two architectures is carried over scenes of varying complexity. While both networks perform well on recognition and bounding box estimation, experimental results show that Mask R-CNN generally outperforms Faster R-CNN, particularly because of the accurate mask prediction generated by this network. These results support well the requirements imposed by an automated task allocation mechanism for a group of unmanned vehicles.