The Study Of Person Tracking Algorithm With A Mobile Robot Based On Multi-Features

With the development of the economy and people's living standard,various of service robots have been developed.Object tracking technology is the key to realize interaction between the service robots and users.Recently,lots of object tracking algorithms have been proposed and achieved good performance.However,most of these methods fail in the object tracking system with a mobile robot,because the robot and people are moving together.Moreover,the changes on illumination and pose,and the occlusion problem restrict the development of the object tracking technology with a mobile robot.With the development of computer hardware and high performance transducers,it is feasible to obtain rich information of the scene.Therefore,more and more researchers have focused on the study of multi-sensors fusion technology and advanced object tracking algorithms.They have made certain achievements.After studying the object tracking algorithms,the advanced object tracking algorithms based on multiple sensors are proposed to realize object tracking with a mobile robot in complex environment.The main contents include the following aspects:(1)double collaboration positioning mechanism based on multi-sensorsObject tracking with a mobile robot is realized by processing the scene information obtained from sensors.However,a single sensor cannot provide enough information for detecting object due to the limited perception ability.Therefore,multi-sensors fusion technology based on RFID and vision is studied to deal with the problems of illumination variations,pose changes,and sudden turning.Furthermore,based on the RFID and vision sensors,double collaboration positioning mechanism is proposed.RFID is employed to locate the object wearing labels,which provides the coarse location.Vision sensor provides image information of the scene.The region of interest(ROI)of the image is determined by the coarse location.Then,the final location of the object is obtained by processing the ROI of the image.The RFID technology narrows the region for image processing,which improves the capability of real-time process.Furthermore,the field of vision(FOV)of the RFID system is larger than that of the vision sensor.Therefore,the robot can adjust its speed to follow the object according to the location obtained from the RFID system,when the object turns suddenly out of the FOV of the camera.(2)object recognition based on gait analysisObject tracking system implements the template initialization task in the stage of initialization.The common used method is to select the object manually,which reduces the flexibility and adaptability of the system.Therefore,object recognitionmethod based on gait analysis is proposed to solve the problem of self-initialization.After processing the gait sequences captured from camera,the LK gait flow image(LKGFI)and head and shoulder mean shape(HSMS)are extracted as the representations of an object when the robot is still and the object is moving.The LKGFI and HSMS capture the motion and static information,respectively.They improve the recognition ability of the gait.Furthermore,the view between the object and robot is estimated to select gait feature in the gait database.The proposed method can deal with the view problem and improve recognition rate.(3)object tracking based on multiple features.The appearance model obtained by using a single feature often fails in object tracking when there are similar interference from the background.The Mean Shift algorithm fusing depth and color information is proposed,and it is implemented under the double collaboration positioning mechanism.The method constructs an adaptive Kernel function by using the depth feature.The kernel function reduces the interference from the background by assigning zero weights to the pixels around the object.Then,multiple features based object tracking algorithm is proposed.Multiple features including depth,shape,color,texture,and motion are extracted for modeling an object.The obtained object appearance model has a powerful resolution ability.In the tracking process,the sequential detection method is used.First,the system benefits from the motion information to narrow the searching region.Then,using the depth and shape information,the head and shoulder model is detected to locate a person.At last,the features including depth,color,texture are extracted for modeling the object.The proposed method has a powerful resolution capability and improves the stability of object tracking.Furthermore,the blocks based multi-features algorithm is proposed for object tracking.The depth image and color image captured from vision sensor are divided into NN ? blocks.Then,the features including depth,color,texture,and motion are extracted for modeling each image patch.The sequential strategy is employed for locating the object as tracking evolves.The depth information is processed for adjusting the rectangle box to represent the object.Then,the occlusion,illumination changes and pose variations are detected and handled by analyzing the depth histograms and the patches' appearance similarity.Finally,an online updating strategy is proposed to deal with the variations on illumination and pose,and occlusion problem.(4)object tracking method based on improved multiple instance learning algorithmThe adaptive appearance model plays an important role in object tracking with a mobile robot.The proposed algorithm is implemented under the double collaborationpositioning mechanism.To overcome the tracking failure due to large motion between two frames,the coarse location obtained by the RFID is used to determine the region for extracting the candidate samples of the IMIL algorithm;To implement real-time tracking,the algorithm improves the MIL algorithm in terms of appearance representation and weak classifier selection.To represent the object accurately,a scale prediction model considering the scales in previous two frames is proposed to adjust the rectangle box online.To deal with the variations on illumination and pose,an updating algorithm based on feedback mechanism is proposed.The updating algorithm changes the parameters of the classifier considering the classification score in the current frame and the error between the classification scores in two successive frames.A fuzzy based intelligent controller is proposed to control the robot.The controller adjusts the linear velocity and turning gain of the robot by considering the relative distance between the human and the object and the object's speed.Using the intelligent controller,the robot adjusts the speeds of the two wheels real time.Finally,the proposed methods are evaluated on a mobile robot.The experimental results show that the multi-sensors can provide more information for the robot,which is useful for improving the performance of object tracking.The gait recognition method adopted in the initialization stage realizes self-starting for the mobile robot.The multi-features based object tracking algorithm represents the object effectively and deals with the occlusion problem in the tracking process.The improved multiple instance learning based object tracking algorithm tracks the object more stably when there are changes on illumination and pose.