Performance Evaluation of Low-Cost Machine Vision Cameras for\n Image-Based Grasp Verification

摘要

Grasp verification is advantageous for autonomous manipulation robots as they\nprovide the feedback required for higher level planning components about\nsuccessful task completion. However, a major obstacle in doing grasp\nverification is sensor selection. In this paper, we propose a vision based\ngrasp verification system using machine vision cameras, with the verification\nproblem formulated as an image classification task. Machine vision cameras\nconsist of a camera and a processing unit capable of on-board deep learning\ninference. The inference in these low-power hardware are done near the data\nsource, reducing the robot's dependence on a centralized server, leading to\nreduced latency, and improved reliability. Machine vision cameras provide the\ndeep learning inference capabilities using different neural accelerators.\nAlthough, it is not clear from the documentation of these cameras what is the\neffect of these neural accelerators on performance metrics such as latency and\nthroughput. To systematically benchmark these machine vision cameras, we\npropose a parameterized model generator that generates end to end models of\nConvolutional Neural Networks(CNN). Using these generated models we benchmark\nlatency and throughput of two machine vision cameras, JeVois A33 and Sipeed\nMaix Bit. Our experiments demonstrate that the selected machine vision camera\nand the deep learning models can robustly verify grasp with 97% per frame\naccuracy.\n