Research On Scene Objection Recognition Methods Based On Machine Vision

Object recognition based mobile robot is a complex subject which combines digital imageprocessing, pattern recognition, robotics theory and many other subjects, which is fully intergratedinto the industrial automation, military, defenseand medical care et al. Since the object recognitionof mobile robot is the basis for indoor and outdoor robot navigation, robot path planning, the robotmap building, etc. Significant progress has been made in Object recognition during the last fewyears. But the environment which is applied to the object recognition for mobile robot is often incase of illumimation, occlusion, and shadow. And the object itself will be recognized variouschanges.According to the classification of objects, object recognition can be divided into dynamic andstatic object recognition.Combining frontier academic knowledgesfor the problems of dtnamicmulti-objects recognition and scene object recognition, and present a new alogorith to improverecogniyion accuracy and robustness. Finally, it’s applied to the robot platform with excellentaccuracy and timeliness.The video and sample images used in this paper which is collected in the Hangzhou DanziUnversity Campus by the mobile robot. The main research work in this paper will be summarizedas follows:(1)Objected segmented is one of the most important steps in the process of dynamic objectsrecognition.Comparing the results of the three object segmention methods: two frame difference,three frame difference, Gaussian mixture model. The pixel of the two adjacent frames is comparedin the two frame segmention. The pixel of the three adjacent frames is compared in the Three framedifference. Gaussian mixture model is based on pixels, Gaussian model is used to calculate thechange of the pixels along the time. Due to the completeness of the segmention dynamic profile andtimeliness of segmention method, two frame difference is chosen to segment the dynamics objects.(2) A real-time dynamic multi-targets recognition method is presented based on machine vision.Firstly, according to pixel changes between the front and rear fames, then segment the movingmulti-targets. Using the Gabor filter to extract features of segmented moving targets and sampleimages to obtain feature vectors. Finally, matching the Fisher criterion classification and marked theresults in the output images. The marked images output continuously，and we can get the outputvideo which has been marked with the results. Experimental results show that: in the case of manydynamic targets moving, the proposed method can effectively recognize. (3) As scene object recognition is one of the important issues in scene understanding, a novelmethod is proposed for scene object recognition based on visual saliency. Inspired by biologicalvisual mechanism, visual saliency mechanism is used to highlight interesting areas in the scene.Firstly, GBVS model is introduced to efficiently screen image data to obtain the salient region ofsignificant interest. Then, GrabCut algorithm based on Graph Cuts Theory is used to extract thesalient goals from salient region. Finally, SURF features are applied to describe the target object,and the bags of visual word for targets are generated through the SURF features learning. Theknowledge mapping between the image features and semantic description of target is achieved bythe SVM classifier matching for bags of visual word. The approach was tested by the image libraryof MIT, and the experiment results demonstrate that the proposed approach has a high rate ofrecognition.