Research On Object Recognition Algorithm For Indoor Mobile Robots Based On Binary Descriptors

With the rapid de elopment of robot technologies,small indoor mobile robots for accompany,housework and the aged assistance ha e become a hot topic in robot fields.As the basis of en ironment perception and accomplishment of complex tasks for mobile robots,real-time and accurate object recognition is significant to research.At present,generic object recognition algorithms are being exploring.It is difficult for classical object recognition algorithms based on local features to achie e balance between real time and accuracy.Besides,indoor complex backgrounds will degrade performance of object recognition algorithms based on binary descriptors.Therefore,a specific research for object recognition algorithms should take hardware of mobile robots and complex situations of indoor scenes into account.This thesis combines an objectness proposal algorithm with an object recognition algorithm based on binary descriptors for indoor mobile robots.The method can pro ide good performance on real time and accuracy.The main contributions of this work are as follows:For eliminating the interference of indoor complex backgrounds with object recognition algorithms,preprocessing of scene images is added before online recognition,based on a research of saliency detection theories in acti e ision fields.The method of the preprocessing step using an objectness proposal algorithm based on a feature called BING,detects candidate object regions in scene images.Then the depth information acquired by Kinect 1.0 and the edge information of scene images are used to select and optimize the candidate object regions.An algorithm called ORB which has good performance on real time and accuracy is chosen by performance e aluation of arious binary algorithms based on local features.Howe er,the ORB algorithm does not ha e scale in ariant and its accuracy on lacking texture objects is not high.To figure out these problems,the modified ORB algorithm detects feature points in scale space and combines with color information to optimize the method of feature description.The modified ORB algorithm is applied as the object recognition algorithm of the online recognition step.It detects,descripts and matches the features of candidate object regions in scene images and object images.The method proposed in this thesis is tested on a self-made indoor mobile robot date set.The experimental results indicate that this method can meet requirements of object recognition tasks for indoor mobile robots on real time and accuracy.Adding the preprocessing of scene images can impro e performance of the object recognition algorithm.