GMM-based visual attention for target selection of indoor robotic tasks
Purpose ‐ Indoor robotic tasks frequently specify objects. For these applications, this paper aims to propose an object-based attention method using task-relevant feature for target selection. The task-relevant feature(s) are deduced from the learned object
representation in semantic memory (SM), and low dimensional bias feature templates are obtained using Gaussian mixture model (GMM) to get an efficient attention process. This method can be used to select target in a scene which forms a task-specific representation of the environment and improves
the scene understanding by driving the robot to a position in which the objects of interest can be detected with a smaller error probability. Design/methodology/approach ‐ Task definition and object representation in SM are proposed, and bias feature templates are
obtained using GMM deduction for features from high dimension to low dimension. Mean shift method is used to segment the visual scene into discrete proto-objects. Given a task-specific object, the top-down bias attention uses obtained statistical knowledge of the visual features of the desired
target to impact proto-objects and generate the saliency map by combining with the bottom-up saliency-based attention so as to maximize target detection speed. Findings ‐ Experimental results show that the proposed GMM-based attention model provides an effective and
efficient method for task-specific target selection under different conditions. The promising results show that the method may provide good approximation to how humans combine target cues to optimize target selection. Practical implications ‐ The present method has
been successfully applied in plenty of natural scenes of indoor robotic tasks. The proposed method has a wide range of applications and is using for an intelligent homecare robot cognitive control project. Due to the computational cost, the current implementation of this method has some limitations
in real-time application. Originality/value ‐ The novel attention model which uses GMM to get the bias feature templates is proposed for attention competition. It provides a solution for object-based attention, and it is effective and efficient to improve search speed
due to the autonomous deduction of features. The proposed model is adaptive without requiring predefined distinct types of features for task-specific objects.
- Access Key
- Free content
- Partial Free content
- New content
- Open access content
- Partial Open access content
- Subscribed content
- Partial Subscribed content
- Free trial content