Object Detection Of Service Robot Based On Deep Learning

In recent years,robots for specific scenarios such as intelligent service robots and elderly service robots have developed rapidly.Scene understanding based on object detection is a very important research topic in the research of such robots.However,in practical application,the complexity of the scene background and the unequal size of the target can easily cause the problem of reduced accuracy of the object detection algorithm.In addition,most of the existing object detection algorithms focus on the improvement of accuracy,which leads to the large amount of model parameters,and cannot meet the requirements of the storage and detection speed of the model applied to the service robot.In response to the above problems,this paper takes the application scene of indoor service robot as the research object,uses deep learning method to study the related object detection algorithm,forms the accurate and real-time semantic understanding of the object category,position and mutual relationship in the indoor scene.The main research contents are as follows.(1)Establishment of complex indoor scene dataset based on Kinect.For application scenarios of indoor service robots,Kinect is used to collect color images of complex indoor scenes under different backgrounds,light intensities,and angles,and the indoor scene datasets are amplified based on data enhancement technology to construct a dataset for object detection in indoor scene.(2)Indoor scene object detection algorithm based on multi-scale feature fusion.Based on the complementarity and correlation of multi-scale features,the SSD object detection model based on multi-scale feature fusion is constructed and the comparative experiments of Faster R-CNN,YOLOv3,SSD and improved SSD are carried out on indoor scene dataset by using the idea of transfer learning to verify the effectiveness of the algorithm for improving the accuracy of multi-scale target detection.(3)Lightweight object detection based on depth separable convolution.The depth separable residual module is proposed to construct a lightweight SSD algorithm based on depth separable convolution.The comparative experiment with the existing lightweight object detection networks Mobile Net-SSD and Tiny-YOLOv3 is conducted on the indoor scene dataset.The experimental results show that the algorithm can greatly reduce the amount of model parameters and improve the detection speed.(4)Real time object detection model for service robots.Combined with multi-scale feature fusion and depth separable convolution,a high-precision and real-time object detection model for service robots is constructed,and the model is tested on the video to verify the real-time and scalability of the model.Through the research on the accuracy and lightweight of the object detection algorithm for service robots in complex indoor scenes,a real-time object detection model for service robots is constructed.The detection accuracy of the model can reach94.66%,the parameter amount is only 40.8MB,and it only takes only 0.027 seconds to detect a picture,which promotes the development of interaction between the robot and the environment.