Target Detection Based on Improved Mask Rcnn in Service Robot

Abstract

Target detection is one of the core algorithms in robot applications, and the recognition speed has a significant impact on robot’s target capture. In this paper, the scene of long-distance and small targets is used as the test scene, and the purpose is to enhance the speed of detection without reducing the accuracy of detection. Consider that the mask branch and excessive full connection layer in the Mask Rcnn network will take up a lot of network detection time, and the feature map extracted by the convolutional neural network has a high dimension, which will occupy a large amount of computational memory. So, in this paper the Mask Rcnn network is improved: remove the mask branch; introduce Light-Head Rcnn into the Mask Rcnn network, increase R-CNN subnet and RoI warping; adjust the proportion of the Anchor in the RPN network. Finally, the improved model is applied to the tensorflow framework. These methods can save computer memory space and improve detection’s speed. In the end, the improved Mask Rcnn network has been verified in a service robot platform with Kinect II. The test results show that compared with Faster Rcnn, the improved Mask Rcnn can have a high accuracy of detection; Compared with the original Mask Rcnn, the improved Mask Rcnn network can greatly improve the speed of the algorithm while ensuring the detection accuracy. The detection time is reduced by more than 2 times, which helps to improve the efficiency of the service robot’s target capture task.