Complex Scene Reconstruction And Target Tracking Based On Global Vision And Self-Vision

With the development of computer vision and the popularization of intelligent monitoring system,3D scene reconstruction,visual object detection and tracking technology are widely used in various fields.However,problems and difficulties still occur when comes to complex scenes,such as limited camera shooting range,distorted image edge,blocked target,light changes and complex background in tracking process,etc.Based on above context,this paper takes intelligent factory,warehouse and logistics as application scenarios to study and improve the stereo matching technology and target tracking algorithm,so as to enhance the reliability and stability of the system.The core content and main innovations are as follows:Based on the study of multi-camera global image mosaics technology.mosaics to the overall flow and structure model is proposed.This paper also compares and analyzes the technologies of image registration and image fusion,which are the key steps of global image mosaics,including Harris corner,SIFT,SURF feature extraction algorithms,mean fusion and dynamic weighted fusion methods.With the calculation of overlapping region in single response matrix transformation,this paper obtains the multi-image mosaic model,and proposes an improved matching method based on overlapping sub-region to solve the problem of poor image registration rate and accuracy.The experiment simulates the environment of a logistics factory and establishes the global vision system,effectively reduces the system complexity of target detection and tracking in the follow-up research.Aiming at the problems of texture shortage,edge depth discontinuity and low matching accuracy of binocular stereo matching algorithm in the scene of logistics warehouse and factory,this paper puts forward siamese convolutional neural network based on the SLIC super-pixel segmentation.The pixel blocks formed by mean filtering preprocessing and super-pixel segmentation are used as the input of siamese convolution network to calculate the local information cost and matching.Meanwhile,the idea of asymmetric design and depth separability is introduced to optimize the convolutional structure of siamese networks.The experimental results show that the matching effect is improved significantly,the average parallax error as well as the average error rate are reduced,the edge region matching accuracy is improved effectively.To solve the real-time update problem in tracking algorithm of mean drift,which is likely to lead to mismatching or tracking loss of the target,a tracking algorithm of mean drift based on deep learning target extraction is proposed.Using YOLO target detection results to initialize the tracking area,iteratively calculate the bhattacharyya distance between the current target model and the next model,determine whether to update the target model,and then according to the target detection results,find the best matching window,effectively improve the performance of the tracking algorithm.Combining with the global vision system,multi-angle and multi-camera image fusion is realized to further improve the success rate of target tracking.The experimental results show that the proposed method can deal with complex and changeable scenes such as background interference,local occlusion,illumination change,and significant change of target size,and has better tracking robustness.