Research Location Research On Object Detection And Location In Shelf Environment On Object Detection And Shelf Environment

As automation can effectively improve the deliverey efficiency and reduce labor cost,the automation of warehouses is the inevitable requirement for the expansion of the business industry.Nowadays,the majority of the tasks in the fullfillment center has achieved complete automation,while the sorting of the shelves has occupied lots of mainpower.The object dection and location algorithm for sorting is still not so mature and stabel that few automatic sorting system has been applied in the warehouse and the sorting automation has been considered as the last mile of the full automation.With the development of electricity business,the requirement for the effective automatic sorting is higher and higher,and the goods detection and location oriented to shelf sorting has been a urgent problem to be solved.In this dissertation,based on the open source operating system ROS platform,deep learning framework Caffe and the pointcloud library PCL,some reseaches about object dection and localtion have been done and the search is specified at the shelf environment.Beside,a shelf detection and localtion system has been constructed with the vision sensor and a Jaco robor arm.Taking the number of goods,lighting conditions and reflective floors into consideration,SSD has been used as the basic network,and the proposed region of SSD has been optimized with by the way learned from the the Faster RCNN and YOLO9000 networks.With the optimal SSD network,the regression of the target region has been more accurate,and the percision of the detection has been improved.Besides,to estimate the pose of target object,the main method of vision measurement has been used to finish the accurate calibration of Xtion.A new approach for pose estimation combined with Nearest Neighbor Search algorithms,Particle Filter algorithm and iterative Closest Points algoriths has been proposed.What's more,in order to solve the problem of partial view and partial depth data,the jaco arm has been used to move the depth sensor and help to extract the feature and pointcloud of the target object from multi views.When controlling the the robot arm,the kinematics model of this robot arm is constructed by DH method,and its inverse kinematics equation is obtained by Newton iteration method.In this dissertation,ROS has been taken as the core to construst the software system to detect and locate targets in the shelf.With the assist of the robot arm,some experiments have been carried and targets have been successfully detected and located.These experiments have acieved good results and the feasiblity and validity of the proposed detection and location methods have been verified.