Heterologous Image Matching Algorithm Based On Deep Learning And Development Of Stereo Vision Positioning System For Livestock Targets

In recent years,driven by big data and artificial intelligence technology,the field of robotics has developed rapidly,and many intelligent robots appear in people’s production and life.The binocular vision system,as the eyes of a mobile robot,is a key component of the robot’s perception and exploration of the unknown world.At the same time,heterogeneous binocular can better capture the complementary information of the environment,and has more advantages in living object detection.Facing the application environment of large-scale farm patrol robots,this paper designs a heterogeneous image feature point matching algorithm and a livestock and poultry target binocular positioning system for the inspection robot.First designed a heterologous applied to the image feature point matching algorithm.Proposed a network based on feature points extracted residuals convolution,the heterologous detection characteristic points in the image;network-based measure is proposed wherein two-channel convolution,a heterologous region of the image comparison between the feature points of similarity.The repetition rate,accuracy rate and recall rate of feature point matching are 33.4%,56.4% and 77.7% respectively,which is greatly improved compared with the traditional algorithm.Secondly,the characteristics of heterogeneous camera imaging spectrometer,designed a heterogeneous binocular camera calibration method.In-depth study of the mathematical model of the camera,analyzed Zhang Zhengyou calibration algorithm theory,completed the design of infrared camera and an RGB camera calibration device and calibration method,calibrated by experimental infrared camera and an RGB camera intrinsic and extrinsic parameters.Calibration errors were 1.15 pixels0.72 pixels.Then,the stereo vision positioning system of livestock target is designed.Using livestock detection data set,the YOLOv4 deep learning target detection model is trained.The data set of feature points matching for livestock is established,and the neural network of heterogeneous image feature points matching is trained again through transfer learning.The matching repetition rate,accuracy rate and recall rate reach 35.1%,58.3% and 79.1% respectively.Then,according to the principle of stereoscopic visual positioning of spatial points to locate matching points,using the center of gravity method to locate livestock targets.Finally,in order to verify the performance indicators of the system,a livestock and poultry target detection experiment,a heterogeneous image feature point matching experiment,and a livestock target binocular visual positioning experiment were successively carried out.The average accuracy of infrared and RGB target detection is 64.35% and 71.31%,respectively.The average positioning error of livestock and poultry targets is 152.7mm.The average processing time of the system is 94.3ms.