Object detection and pose estimation is essential in many applications. Information about pose is necessary for robotics (industrial and service), augmented reality, and automatic visual inspection. Research on pose estimation methods has also opened up many new areas of estimation utilization. A common application in industrial environment is bin-picking, where a robotic arm picks and orders cluttered objects. In recent years, many researchers have focused on service robotics, where a robot solves tasks in domestic environments; for example, stacking dishes in the dishwasher or cleaning. In addition to robotics, a known pose is also important for automatic visual inspection. Visual inspection systems are often crucial for the detection of defects in manufactured products and for diagnosing problems in a manufacturing process. Currently, one characteristic of automated visual inspection systems is their specialization. With few exceptions, nearly all of the existing automated visual inspection systems have been designed to inspect a single object or a part of one whose position is highly constrained. Positioning is usually achieved by mechanical manipulation of an object, a process that can be expensive, space and/or time consuming, or simply impossible in some scenarios. A visual inspection system that could inspect arbitrarily positioned objects would bypass the need for the mechanical manipulation of inspected parts, thus reducing the cost of the system and increasing its exibility. It would also enable inspection in scenarios where previously it would have been thought to be unfeasible. For visual inspection in mechanically unconstrained environments, the system needs to detect and estimate the pose of the objects, before the inspection can take place. Pose is important for visual inspection in ordered environments where we can constrain (partly) an object's pose by mechanical manipulation and particularly in cluttered environments where the pose is unknown. The original contributions of this doctoral thesis include the development and evaluation of methods for the pose estimation of multiple identical objects in heavily cluttered environments as also a method for in-plane rotation estimation. Original contributions are briey presented below.
|
