Exploiting Parallelism for Object Recognition on Mobile Robots

Object recognition is a key problem in robotics research. When this process is executed sequentially as is typically done, it can take a fair amount of time (several minutes). However, if we deploy such a system in the real world for robots operating in homes and offices, we would ideally expect this task to be done in close to real-time (less than ten seconds). In this project, we plan to utilize computing resources around us (e.g. Emulab [1]) to speed up object recognition. Segmentation is one of the most time consuming, complex and challenging parts of object recognition. It involves splitting sensor data, such as a 3D laser scan or camera image, into constituent pieces which serve as possible hypotheses of objects to be recognized (Fig. 1 is an example segmentation). We focus on speeding up segmentation in this project. Segmentation is difficult to parallelize because there is no natural way to subdivide the segmentation task. This is unlike tasks that can be decomposed into independent subtasks (e.g. MapReduce[4] tasks). Also, segmentation is not easily parallelizable because of data dependencies; there is information that needs to be shared among all subtasks. Given this, it is particularly difficult to effectively speed up segmentation. We believe our investigation can yield useful preliminary findings and lessons for others wishing to build similar systems. Our main contributions are as follows: