Obstacle Detection Of Indoor Service Robot Based On Image Understanding

In the application of indoor service robot,obstacle detection is a very important part of indoor service.Because the vision of indoor service robot is different from our artificial vision,this paper starts from the perspective of indoor service robot,and makes a new understanding of the obstacles in its vision,and builds a data set for obstacle detection.And the specific research contents are as follows:The common indoor positioning technology mainly includes ultrasonic,Bluetooth,WiFi and visual positioning,among which the image-based indoor positioning technology has high accuracy.Considering the accuracy and stability of different cameras and the consumption of computer memory,this paper selects the image sensor based on depth camera for indoor obstacle detection.In this paper,we select 10 kinds of targets on the open data set VOC 2007,and test the accuracy and real-time performance of three algorithms,Faster R-CNN?YOLO,YOLO algorithm and SSD algorithm.The test results show that Faster R-CNN has the highest recognition rate,but its processing speed is very slow.The real-time performance of YOLO and SSD is similar,and the overall detection effect of SSD is better.So this paper chooses to use SSD algorithm to realize obstacle detection and improve it.In order to improve the real-time performance of the deep learning algorithm,a new multiscale convolutional neural network is constructed by combining SSD algorithm with Mobilenet.From the perspective of the robot,the data set can be divided into 10 categories: bucket,flower track,garbage basket,soft leg,box,chair,computer,fire extinguisher and trigger.The comparison between the original algorithm and the improved algorithm in the open data set VOC 2007 and the self built data set shows that although the improved algorithm slightly reduces the accuracy of target recognition,its processing speed has been greatly improved,which proves the reliability of the improved algorithm in this paper.In this paper,Kinect depth camera is used to measure the depth information of obstacles.Through deep learning,the obstacle is identified to get the image segmentation area,and the positive midpoint of the area is found.The depth information of the positive midpoint is set as the depth information of the obstacle target,and the depth information measurement experiment is carried out.The experimental results meet the initial requirements.Finally,based on the back-end robot obstacle detection platform,the front-end robot collects image information,and the back-end PC completes the implementation of the core algorithm and the upper computer display.Using 10 kinds of obstacles in the self built data set to measure the accuracy of depth information,the measurement results show that the closer the distance is,the higher the accuracy of recognition of depth information of obstacles is,and the average distance accuracy can reach 0.932 when the distance is set to 60 cm,which can meet the equal accuracy requirements of indoor service robots for obstacle avoidance.Through the experiments and analysis of this paper,we can prove the effectiveness of the improved deep learning algorithm and the stability and reliability of the obstacle detection of the whole indoor service robot.