Research On Indoor Object Recognition Algorithm Based On Object Candidate Regions And Improved Random Ferns

With the development of science and technology,robot technology has been rapidly developed.Indoor mobile robots are attracting more and more attention because of its extensive application in family services,medical rehabilitation and other important areas.In order to complete a series of complex tasks in the practical application and achieve its autonomous intelligent,the research of mobile robot's visual perception ability is particularly important.As the basic problem of visual perception,object recognition has become the basic task of indoor mobile robots.However,object recognition algorithm in the traditional field of computer vision is facing many difficulties.For mobile robots in a complex indoor scene,more challenges will be faced.Therefore,it is of great significance to find a fast and accurate object recognition algorithm which is suitable for indoor mobile robots.Based on the research of salience detection in the active vision theory,and the wide application of depth information in the field of computer vision in recent years,a method of object candidate regions detection is proposed combining BING(Binarized Normed Gradients)features and scene depth information.Firstly,some object candidate windows are quickly obtained using BING features,and then the large candidate set is screened using the depth information,leaving a small amount of candidate windows with the highest likelihood.Finally,taking the recognition accuracy and efficiency into account,the object candidate window set is extended and grouped.After the object candidate regions are obtained,the target object recognition is carried out by the random ferns algorithm with the characteristics of online learning and human assistance.When the target object is partially occluded,the recognition performance of the original random ferns algorithm is not good.So in this paper,a “sub-regions” concept is proposed to improve the original algorithm.The improved algorithm is more robust to object recognition with partial occlusion.Meanwhile,the method of obtaining the fern features by comparing the intensity values of two pixels directly to generate the binary string in the original algorithm is modified,resulting in an improvement in the recognition performance for the target object with less texture.Four representative objects are chosen for an experiment to validate the good recognition performance of the improved random ferns algorithm.In the end,a data set is built for real indoor scene.The object recognition algorithm in this paper is evaluated in this data set.And the result shows that the recognition algorithm in this paper has a good recognition performance on the basis of meeting the real-time requirement of mobile robots.In an addition,according to comparing the recognition results before and after the object candidate regions detection is added,it can be verified that the object candidate regions detection has great help for improving the recognition accuracy.