Automatic Training Method of Deep Neural Network for Robot Vision

Abstract

Listen

Translate article

Thanks to the breakthrough of deep learning in machine vision, robots have been widely applied in industrial and family services in recent years. Object capture is one of the most important functions for robots, and two-dimensional object detection is the premise for robot to capture objects. However, the training cost of obtaining the high recognition rate model is very large, as the deep neural network needs huge data samples. Based on this, this paper proposes an automatic training method of deep neural network for robot vision. The Tracking-Learning-Detection (TLD) algorithm tracks and collects the object samples by online learning. Then the offline Single-Shot-Detector (SSD) model studies the features of the object so as to realize the robotic object recognition function. In order to ensure the stability and continuity of the tracking process of the target, the sample acquisition process is accomplished automatically by the robot manipulator. Two methods of manual annotation and TLD algorithm are all used in this paper for increasing the persuasiveness of the data. The results show that the time cost of automatic data annotation by TLD algorithm is 77.75% less than the manual data annotation, and the recognition rate of model trained with automatic labeling data is 97.75%, which verify the validity and feasibility of the novel method.KeywordsRobot armOnline learningOffline learningAutomatic data annotation