Research On Visual Cognitive Technology In Robot Mobile Operation

With the advent of the artificial intelligence era,robots are increasingly used in a wide range of industries.In the case of indoor construction sites,robots can reduce labor and save a lot of time.Vision can provide a large amount of information for robots.It is of great significance to study vision-based robot object recognition and positioning systems.This paper aims to imitate the human visual cognition mechanism and give the robot the visual cognition ability of target detection and positioning.The traditional depth camera based on binocular vision is very very sensitive to ambient light.The effect of binocular vision algorithms is drastically reduced in the case of strong and dark illumination.At the same time,in the scene of monotonous lack of texture and lack of visual features,the binocular vision algorithm may also have difficulty in feature matching.In addition,the depth camera based on the time-of-flight method has problems of large power consumption,high cost,and low depth image resolution due to immature technology.Therefore,this article uses a binocular structure light camera to complete the ranging function.In recent years,deep learning has developed rapidly in the field of computer vision.The single-task network structure has gradually become less noticeable.Instead,it is an integrated and complex multi-tasking network model.The representative is the instance segmentation model.Instance segmentation is a more comprehensive problem that combines target detection,image segmentation,and image classification.Therefore,in order to make the robot better and faster to identify the object and determine the position of the object,this paper proposes a method based on Mask R-CNN model and structured light camera for object detection and localization.The main research work of this paper is as follows:1.Research and implement the Mask R-CNN model,and optimize the performance of the model by adjusting the training method of the Mask R-CNN model.One is to ignore the target frame provided in the training set.Instead,use the method of generating the target frame yourself,select the smallest box containing all the pixels of the target object as the bounding box;the second is to use the gradient cutting method to prevent the gradient explosion.Thus the effect of the Mask R-CNN model is improved.So far,this paper uses the Mask R-CNN model to achieve accurate image detection and image segmentation.2.Research depth camera ranging technology and combine depth learning to improve range measurement.The current mainstream method of measuring objects with a depth camera is to average the depth values of each pixel in the entire target frame.However,not every pixel in the target frame belongs to the target object,so this method will inevitably bring errors.In this paper,the Mask R-CNN model is used to perform pixel-level segmentation on objects in the target frame,thereby filtering out most of the pixels that are not targeted,thereby improving the ranging effect.So far,the target detection and positioning functions have been completed using the Mask R-CNN model combined with a binocular structure light camera.3.Finally,based on the target detection and positioning results,the target capture experiment was carried out by the Ur10 robot arm of the Ur company in Denmark.The final experiment proved that the effectiveness of the indoor detection and localization method based on Mask R-CNN model and structured light camera proposed in this paper.