Research On Object Detection,Tracking And Feature Classification Based On Computer Vision

With the rapid development of the Internet of things technology,the video surveillance camera which is an important part of the Internet of things has been deployed in all corners of the world.Especially under the policy that government promote the open block environment,the number of video surveillance cameras will be more and more and the coverage will be greater and greater.However,due to human physiological weakness,video image data cannot be observed in real time.And its more difficult to real-time identify key object in video image.The misaction in the saved video data analysis is easy to be neglected.The application of computer vision technology to analysis and understanding the object action in massive video image data has become an increasingly urgent demand.Object action recognition in intelligent monitoring system will have great research value.In this paper,based on the analysis of video image data,using computer vision,we detect and extract interesting objects,segment the effective information from the complex background,analyze the action of the object,and understand the semantic of the object action under the context.The research results can be applied to intelligent monitoring,abnormal early warning,motion analysis,video retrieval,human-computer interaction and other fields directly.In recent years,the field of object action recognition has attracted a large number of researchers to participate in the study.The object action recognition is an extremely popular direction in theory research and practical application.At present,a large number of researches have been done on the object action recognition of simple background in the experimental environment.Yet there is no detailed research to the analysis of object action in complex environment and complex action.Action analysis and object action understanding technology of video still exist the following problems: 1)the object recognition in complex scenes.There are large amount of noise in the complex environment,it has a great influence on the object action analysis.Therefore,the analysis of human action in complex environment is more challenging.How to filter out a large number of noise in the complex environment and improve the accuracy of action recognition is the focus of the current research object action analysis.2 fast object action detection and recognition.Action detection and recognition is actually finding the spatial location of the specific action in the search video and its content semantics.Usually,the specific action in time and space scales is unknown and the meaning of action is different.The video image with the iterative method of exhaustive search will tend to increase the computational time and complexity.Therefore,in order to save the computation time and complexity while guaranteeing real-time detection and precision of the action analysis,we need to do further investigation.Fast object detection and recognition is a difficult problem in the action object action analysis.Aiming at the problem of object action understanding based on computer vision in the Internet of things video sensor,we extend our work from the four stages of action recognition: object detection,object tracking,object feature extraction and action analysis.In the analysis of the video object action,we combine with computer vision and image processing methods.The key technologies such as feature extraction and description,object action modeling and object action recognition in video surveillance scenes are deeply studied.From the aspects of robustness and computational efficiency,the key problems of the action recognition method are studied.Main research work and innovation are as follows:(1)Aiming at the traditional MRF models only consider the characteristics and its neighborhood spatial correlation,this paper presents a new video dynamic object detection algorithm base on space-time Markov Gauss mixture model using random field(GMM-STMRF).Firstly,the Markov random field model is extended to increase the correlation of adjacent pixels.In the training process of the Markov random field(MRF),the mean value and variance of each region in time domain are calculated by using the parameter updating method of Gauss mixture model.In addition,the energy function calculation method of space-time Markov random field is further improved,and the value of the potential field in the neighborhood of time and space is considered.Comparing the proposed algorithm with the standard Gauss mixture model,Meanshift,FCM in the performance,this paper proves that the algorithm has a good effect on the object detection and extraction in the video and the complex environment.What's more,it has better detection accuracy,robustness and time complexity.(2)Because intelligent video surveillance images is of poor quality and contains a great number of noises,object detection is not accurate.We proposed a fuzzy object tracking method based on improved particle filter(Particle Filter,PF).In the framework of the standard particle filter,a nonlinear non Gauss multi state space fusion model is constructed based on frame difference of video multi image.The particle sampling and the spread of probability density is carried on in the range of the key points of the frame difference.The posterior probability density of the state space fusion model of multi system is presented with weighted posterior sample particle.System posterior state is estimated using sample mean method.Finally,the state space model of the system is output to get the object tracking results.The experiments are carried out by using the video surveillance database,and the improved particle filter,the standard particle filter,the Calman filter and the extended Calman filter are compared.The experimental results show that the improved multi state space model and key areas sampling particle filter algorithm has a good tracking effect in the fuzzy object tracking problem of video,and it has high time efficiency.(3)Convolutional neural network(CNN)is a kind of deep learning model,but it only extracts visual features.This paper expands the convolutional neural network and studies the extraction algorithm of local space-time features,presents a fusion of spatio-temporal interest point convolution neural network model.The local spatio-temporal feature points and common the visual features are combined to obtain information from the image feature dimension.At the same time,the neural network model is improved,and the convolution operation is performed by the gradient back propagation and the weighted kernel function.The feature extraction and calculation of the model will get the local spatial and temporal motion information of the object.Then the object action can be classified more accurately.The experimental results show that the proposed convolution neural network with the feature points of temporal and spatial features has better learning ability and classification ability.In addition,according to the high dimension and multi feature of convolutional neural network,a method of feature classification based on multi instance learning(MIL)and SVM is proposed.The high dimensional feature clustering is regarded as multi instance learning package.The information annotation package is determined by calculating the diversity characteristics of density function value to get the positive examples set.Finally,the SVM algorithm is extended to a multi sample nonlinear classifier,and the classification and recognition of the sample package is carried out.After several iterations,the classification accuracy and recognition effect are obtained.(4)Gauss mixture model is a common method of object detection and has a better detection results in the fixed scene.But the GMM algorithm models all pixels in image to establish the Gauss model.The time complexity is high.In the high resolution video frame with more pixels,detection result in real time is poor.So it cannot be applied in HD video capture device.According to the background of GMM differential detection algorithm to detect the problem of low efficiency,this paper introduces the membrane parallel computing system which developed based on computational biology theory.We use the task decomposition of GMM algorithm and the maximum parallelism of the membrane computing system.This paper presents a fast object detection method based on membrane computing system for Gauss mixture mode.The test results are as good as the Gauss mixture model.But the speed of the algorithm is greatly improved.Membrane computing model is introduced into the GMM object detection model.GMM training and real-time object detection two stage membrane structure model is constructed.The Gauss background computation process of image pixels is described by means of membrane objects and rules.The video frame image is input into the membrane system environment.The image pixels are divided into row vectors to be transferred into the inner membrane.In the inner membrane,each pixel in the row vector is further transferred into the basic membrane to be calculated in parallel.And a plurality of basic membranes is simultaneously generated by the membrane splitting rule.Pixel parallel processing is implemented by using multiple computing units.Pixel membrane objects is converted into a computing membrane with different functions according to the rules provided by membrane system.We update calculation method of the model parameters for the design of GMM algorithm based on evolution rules.The selection rules and cross rules to complete multi Gauss model sorting and matching threshold update.