The Study On Adaptive Detection Technology Of Moving Target

Mankind has entered the artificial intelligence society in 21^st century,the emergence of robots,speech and image recognition technology facilitate people's work and lives,liberate people from heavy and repetitive jobs.So far,the artificial intelligence technique has been applied to many fields,but the depth of application is still shallow,the combination of technique and application has a long way to go,on the other hand,the knowledge of artificial intelligence need to be completed by academic field.This paper mainly focuses on two aspects:pedestrian detection and three dimensional reconstruction.Pedestrian detection is the core problem of Intelligent Safe-guard System,the detection and location of people is the basis of behavior recognition;the 3d reconstruction tech aims to acquire 3d information,to understand the scene,and is the foundation of virtual reality and augmented reality.For target detection,this paper first to detect object based on traditional background–forehead methods,but this kind of method has the shortcomings such as easily affected by noise,information loss and cannot locate the position of the target.The paper also uses the classic features histogram of oriented gradient and local binary pattern,associated with svm and ensemble classifiers to classify the target and background.Experiment shows that hog feature played a better role in detect the target,but lbp is not a better feature to describe the body of people.In the deep learning,this paper uses cnn to test the local samples based on deep learning frame,but the result is not very well,the final result has great correlation with the distribution of sample dataset.The traditional background–object detection method cannot locate the target,and feature-target detection may false detection or miss the target,and is time-consuming.Thus,a new method is proposed in this paper,this method first use gmm to build the background to extract the coordinate position of the foreground,then using k-mean cluster method to acquire the center of the target.This method could processing the image without delay,and all the target can be found,the detect speed is fast and could play a better role in filter noise,the experiment result is well.But it still has the problem that the number of objects detected by this method is more than the actual.To optimize the number of target,the number of cluster center is determine by floor and ceiling number of target.The combination of cluster detection and feature-classifier reduce the region that needed to be detected,this paper also analysis the time consumption,advantages and disadvantages among feature-classifier,classifier detection,the combination of background and feature.For 3d reconstruction,this paper calibrates single camera and binocular camera by using chessboard calibration method.The data set is based on open source computer vision library,and 3D scene is created based on the images captured by two cameras.The 3D information is obtained from different viewpoints.