ABSTRACT
Possible situations in which the robots are likely to make vision-related misjudgments are discussed, e.g., when there is a lack of knowledge about the precise position or orientation of the robot, the collected data is influenced by the motion of the robot or affected by environmental parameters or when the object of interest is occluded and not visible due to obstruction. Several techniques are described that aim to address the challenges of robot vision. In active vision, the robot can move its sensors to gather more useful information about its surroundings. Use of model- and data-based methods for anomaly detection is explained. Anomalies in the kinematic or dynamic behavior of a robot are detected by comparing its observed motion with the expected motion. During the examination of the vast amount of data, image-of-interest detection is employed as a time-saving measure, where visual representations are provided with specific regions, such as corners and edges, highlighted. In semantic vision, the relationships between objects in an image are understood. Visual place recognition enables the robot to recognize a place based on its visual features, including color and shape. Scale-invariant feature transformation is an algorithm that detects, describes, and matches local features in images. A convolutional neural network-based approach is also useful for this purpose. Simultaneous localization and mapping (SLAM) is a technology that enables robots to build maps of their environments and use these maps to navigate while maintaining their precise location. Vision-based scene understanding analyzes visual data to interpret and derive meaningful information about a scene, e.g., the objects, their relationships, spatial layout, and context. Point clouds (3D representation of a scene with each point representing a specific location in space), depth maps (2D image with each pixel representing the distance of the corresponding point in the scene from the camera), and stereo vision (using two cameras for depth information) are the strategies adopted to help the robot in 3D object detection. All these methods are explained. A comparison of 2D and 3D object detection is made.
