Research On Object Recognition And Localization Based On Convolution Neural Network And Binocular Vision

With the development of the industrial of automation and the computer vision technology,object recognition and location based on binocular vision has become a hot topic in the research of artificial intelligence,nowadays.Object recognition and positioning have broad prospects and huge value in many applications.The technology can be used in self-driving,drones,industrial robots,service robots,military reconnaissance,bionic robots,and planetary probe vehicles.This paper uses the methods of convolutional neural network and binocular vision to achieve object recognition and positioning.The dissertation is focus on theoretical research of binocular vision,object recognition algorithm,image feature point matching algorithm and their improved algorithms.The existing binocular object recognition and positioning technology are based on the algorithm that Combined of artificial design features and machine learning.That method has the disadvantages of single object recognition and low recognition accuracy.This paper put forward a method that combined binocular vision and convolution neural network to realize object recognition and positioning.This method is well extensible which can train the data based on project requirements.In this paper,the SSD and YOLO convolutional neural network models are adopted.The SSD model has the advantages of high detection rate and high recognition accuracy.The YOLO model doesn't rely on any deep learning framework,and it can achieve the real-time detection rate,which can ensure it be well migrated to the embedded Platform.Experimental results show that the method has the advantages of multiple types of object recognition,higher recognition and object positioning accuracy.In practical application,there is no rotation and scale change in binocular images.The traditional feature point matching algorithm integrates the image pyramid structure algorithm,main direction calculation and corner point feature detection method.As a result,the algorithm has some flaws in large calculation amount and micro-texture target detection feature point.To solve these problems,this paper proposes an C-SURF algorithm based on edge feature that adopts the Canny edge detection algorithm to achieve feature point extraction.The improved SURF feature point description algorithm based on the absence of rotation and scale changes between binocular images,improves the computational efficiency.A multi-feature point fusion matching(MPFM)algorithm was proposed.The feature of corner points was added on the basis of the edge feature points to make the feature points evenly distributed,thereby increasing the local position information of the target and solve the problem of the uneven distribution of feature points.The experimental results show that the improved algorithms can solve the problem of micro-texture detection which has few feature points,computational algorithm time-consuming,and uneven distribution of target feature points.The experimental results show that the proposed methods of designing object recognition and positioning are reasonable and feasible.The object recognition and positioning system designed in this paper has certain accuracy and stability.The system is programmed on the pc and embedded platform respectively.The proposed binocular image feature point matching algorithm has certain values in theory and application.