tical flow of pixels. Section 2 discusses the general architecture of a computational level of the system and defines the functions and the interfaces of the task decomposition, world model, and sensory processing modules. Section 3 describes the functions and interfaces specific to Level 1 processing for a camera. Section 4 provides an example of the interactions between modules in performing a typical telerobotic task. Appendix A describes preprocessing algorithms that can be applied at Level 1. Appendix B describes edge point extraction algorithms, surface patch or region extraction algorithms, and the first level of optical flow extraction algorithms. 2. General System Architecture Before describing the functionality of Level 1 of the perception system, a description of the general structure of a computational level is presented. Each level consists of a task decomposition module, a world model support module, and a sensory processing module (fig. 4). The task decomposition module bases its decisions on information extracted by the sensory processing module. The sensory processing module is driven by predictions of the state of the world provided by the world model. The world model maintains the best estimate of the past, current and possible future states of the world [ALBUS81]. 2.1. Task Decomposition The task decomposition module consists of three submodules: Job Assignment (JA), Planner (PL), and Execution (EX). These modules have the same general functions at each level of the system. The Job Assignment module accepts and queues commands from the world modeling support module or the operator. The commands are passed to the Planner module, which analyzes the request and selects the most appropriate sensory processing algorithm for achieving the desired output. The Execution module obtains confidence factors from the world model, updates and modifies algorithm parameters as required, and passes this information through the world model to the sensory processing system. In this way, the evaluation of sensory processing algorithms serves as a learning tool for improving the performance of the algorithm. It is also responsible for activating or deactivating the sensor itself. Each of the three modules execute cyclically to process commands and pass information. They read inputs, perform computations, and generate outputs independent of the other modules. This type of processing allows the system to operate quickly and efficiently. It prevents system deadlock that can occur when one process waits indefinitely on another for data. It also allows the system to respond to new information without being explicitly commanded for updated calculations. To coordinate the requested commands among modules, the Planner and Execution modules are directed by one Job Assignment module. The single Job Assignment module interacts with s Planner modules, where s is the number of classes of processing algorithms at a given level of the system. At Level 1, there are five classes of algorithms: filtering, enhancing, edge point extraction, surface patch extraction, and optical flow. Each of the Planner modules communicates with t Execution modules, where t represents the number of algorithms that supply the type of features in the class (fig. 4). |