Research On Robotic Visual Localization Method For Grasping Weak Luminance Targets

In vision-based grasping,localization is to provide enough grasping information for robots,including position and attitude information of the target,graspable position of the target and grasping posture of the gripper at this graspable position.Based on these information,robots will accomplish other grasping steps and finally perform the grasping operation.When the robot grasping targets based on vision,it is difficult to acquire enough grasping information from the weak luminance image with uneven brightness,blur,and poor contrast.As a consequence,how to obtain the adequate grasping information in weak luminance images to quickly and accurately locate the grasping target,poses a challenge to the research on visual localization of robots.In this work,robotic visual localization methods are studied for grasping weak luminance targets.To acquire the clearness image information and provide the accurate visual information for robots during grasping,this work initially for the issue of improving the obtained image quality,investigates the image contrast enhancement method.Region of interest is introduced,and confirmed via grayscale transformation.As well as,in accordance with global histogram equalization,a contrast enhancement algorithm for region of interest is proposed.The algorithm selectively improves the contrast of the region of interest in the image,with advantages of good contrast enhancement effect on the image,and less computation time.To achieve stable positioning of the target during grasping,in accordance with local invariant features,a pose estimation method based on tracking point-line features is proposed.The method could satisfy the requirements of the grasping speed in robotic grasping,and could realize the accurate estimation of the target pose.Obtaining an accurate initial pose is the premise of the successful implementation of this method.Therefore,an initial pose solving algorithm based on multi-point matching is proposed,which takes the measured pose as the initial value of target tracking.For the sake of adapting to complex and changable environments,during tracking the target,a pose tracking algorithm integrates the point-line feature is proposed,and the M estimator is used to resist the interference of the external environment.When the robot via vision locates the target,it should accurately know the position of the target,but also be able to clearly judge the graspable position of the target.To solve this issue,a real-time grasp detection method based on YOLO v2 network is put forward.The core of this method is to propose an "end-to-end" real-time grasping detection network model with images as input and grasping parameters as output.As well as,to training the generalization ability of the network model,definition of loss function,extend the dataset and transfer learning are conducted.The model could realize the real-time detection of the graspable area of the target and the grasping pose of the gripper,simultanelous maintaing a highly detection accuracy.To verify the methods studied in this work whether could achieve stable localization of the target,thereby to achieve successful grasping,a vision-based robotic grasping experimental platform composed of a visual camera,a robotic arm and computers is established,and vision-based robotic grasping experiments are carried out.Before implement the grasping experiments,to acquire the image information conforms to the real situation,a self-calibration method based on unsupervised vanishing points is adopt to calibrate the internal parameters of the camera.In this work,methods on image contrast enhancement,pose estimation and grasp detection are researched,to promote the localization ability of the robot in vision-based grasping,thereby improving the successful rate of vision-based grasping,and laying a foundation for the application of grasping weak luminance targets.