Font Size: a A A

A Research Of 3D Motion Sensing System Based On Speckles

Posted on:2016-04-03Degree:MasterType:Thesis
Country:ChinaCandidate:S Z LiuFull Text:PDF
GTID:2348330536987014Subject:Control Science and Engineering
Abstract/Summary:PDF Full Text Request
In recent years,how to achieve free and easy interaction between humans and computers is becoming a hotspot scholar,especially with the develop ment of artificial intelligence,machine vision and related hardware.Somatosensory interactive technology as an important research direction of human-computer interaction cannot in the case of the man-machine contact,body movements by people naturally interact with the machine complete information.Aiming somatosensory interactive technology based on optical sensing of two main components,namely the three-dimensional reconstruction and human pose estimation has been the focus of research,in order to achieve a real-time accurate somatosensory interactive method,main tasks are completed:In the 3D scene reconstruction,this paper propose speckle-based 3D scene reconstruction algorithm.Firstly we uses zero-mean normalized cross correlation operator(ZNCC)in 3D scene reconstruction to obtain disparity.And then we use triangulation theory to obtain depth.Since the somatosensory interactive algorithm has high requirements to accuracy and timeliness of scene reconstruction,in order to resolve disparity discontinuous region mismatch problem this paper adopt "winner takes all" and seed growth strategy to solve continuous and discontinuous regions and use a ZNCC relevant operator to improve the matching calculation of noise immunity and adaptive lighting.Finally,in order to resolve the ZNCC operator redundancy calculation,we use integral image and prefix adder accelerate the process of obtaining a disparity based on GPU parallel computing architecture.In the human body posture estimation,we takes attitude estimation as classification of depth pixels based on depth data and completes classifier training of gesture by Random forest algorithm and extracts the skeletal point by mean shift algorithm.In the training process,this paper takes feature based on pre-screening methods to remove the invalid feature.So as to improve body posture forecast accuracy,we lead into hierarchical decision thoughts according to features in the combined action posture in the decision tree.At last,we do some experiments to verify the validity of the proposed method.
Keywords/Search Tags:Speckles, ZNCC, GPU, Hierarchy random forest, Feature Pre-Selection
PDF Full Text Request
Related items