Game Robot's Vision Based on Faster RCNN

Abstract

The traditional object detection algorithms are slow in speed and low in accuracy. With the popularity of the deep learning algorithm, the object detection algorithm based on deep learning has been greatly improved compared with the traditional algorithm. So this paper is based on an improved Faster RCNN which is an excellent object detection algorithm framework to realize game robot's vision. In this algorithm, ResNet101 network extract features and position label mapping maps a single object with it's label. The information of moving chessman is obtained by tracking the change of frame from robot's camera to improve the accuracy and speed of the model recognition. This paper presents a more robust result than the traditional algorithm(canny detection or hough transform) by comparing the results of the single target and the multiple target under poor illumination condition. Our model was trained under small datasets and overcame traditional algorithm's drawback of illumination sensitivity. At the same time, some pruning optimization methods were used to boost detection speed.