Hybrid Saliency-Based Visual Perception Model for Humanoid Robots

Abstract

Recent years have seen an explosion of research on saliency detection and visual attention model development for humanoid robots. The bottom-up and top-down visual saliency detection models can be combined to develop hybrid visual attention for the interaction of the robot with humans. Most of the hybrid visual saliency models are not computationally economical for practical implementation on humanoid robots due to the high computation cost and complexity of their model. The main drawback of most of the visual attention models is that they can detect the salient object in the case of simple background in natural images, but cannot perform well in the case of images having cluttered and textured backgrounds. Most global contrast-based methods do not produce efficient results for images having multiple salient objects. The hybrid models based on local and global contrast-based methods have an issue that most background regions are predicted as salient regions. In this research paper, a hybrid stereo saliency model based on PC/BC-DIM neural network is presented which can efficiently detect salient objects for the simple, cluttered, and textured background images. The proposed model has added advantages of its simplicity, robustness, and execution on a CPU system due to which it is perfectly suited for realization on humanoid robots. The proposed saliency detection model can detect multiple salient objects. The mean absolute error (MAE) score for the hybrid saliency model is 0.22 and for the stereo saliency model, the MAE score is 0.375. The proposed model is computationally efficient and cost effective models for implementation on humanoid robots.