Research On Scene Object Detection And Object Classification Based On Machine Vision

Target detection and classification are hotspots in machine vision.Target detection,classification and semantic analysis are the prerequisites for scene understanding.Main information sources of scene understanding are images and videos.According to these sources and different targets which need to be detected,scene target detection is classified into that based on video sequence and single frame image.The former detects dynamic areas(moving objects)with information about multiple-frame images,then classifies and tracks objects while the latter identifies and extracts all interested objects which are learned in a single image.Correct identification and accurate positioning in the original image are fundamental steps of high-level visual analysis.In recent years,the Convolutional Neural Network(CNN)is an effective method to solve various problems of computer vision.It can acquire features of objects independently by learning large data samples,which avoids complicated process of feature extraction and data reconstruction in traditional recognition algorithm.According to the information above,the main researches in this thesis will be summarized as follows:At first,the method of dynamic target classification is put forward based on CNN.With logic analysis,it solves the problem of multi-target fusion in moving regions extracted by Gaussian Mixture Model(GMM)and acquires comparatively complete and independent moving objects.Then it learns inherent static characteristics of moving objects by the deep convolutional neural network and classifies with softmax regression classifier.The experiment shows that the method can still accurately classify in cases of incomplete detection of moving object region and overlap of same category,and has the merit of high classification accuracy and rapid processing speed.Secondly,a detection method of image targets is proposed based on Aggregated Channel Features(ACF)and CNN.In order to solve the problem of partial false detection when pedestrian and vehicles are detected by ACF,CNN is used to secondly identify the candidate box to get rid of the background box of false detection.The method makes full use of high recall rate of ACF and high recognition rate of CNN.The experiment finally shows that the modified method can improve detection accuracy but not influence real-time and recall rate.Finally,a detection approach of image targets is proposed based on ACF and CNN for multitask learning.This method can not only detect image target fast and accurately,but also increase rich semantic information for targets according to the characteristic of muti-task learning.The experiment of pedestrian and vehicles detection demonstrates that the method can improve the accuracy and get the hybrid semantics of targets of pedestrian and vehicles.It can detect pedestrian in a candidate box,and further detect the direction and state of motion at the same time.The method can be applied to the platform of a mobile robot to analyze abnormal behavior with image position information of several pedestrian and vehicles in scene and mixing behavior of semantic.