Research On Object Recognition And Robotic Grasp Detection Method Based On Deep Learning

Autonomous perception of the surrounding environment is an important prerequisite for robots to complete tasks.Due to the complexity of the real environment and the variety of tasks,robots need different types of sensors to capture information about the surrounding environment.Traditional object recognition methods rely on hand design features,which can not adapt to the complex working environment.Therefore,how to effectively identify the target object becomes a challenge for robots.In addition,the grasp detection methods required by the robot to complete the task also rely on the known information of the target object,which cannot effectively deal with the new and complex environment.This dissertation proposes a target recognition algorithm based on deep learning,which enables the robot to accurately identify target objects in two-dimensional and three-dimensional spaces.Moreover,a deep learning based grasp detection method is further proposed,which enables the robot to autonomously learn the grasp positions,and improves the adaptability of the robot to the new environment.Aiming at the recognition of occluded objects in two-dimensional image space,an object detection post-processing method based on improved harmony search algorithm is proposed.A combined optimization method is used instead of the non-maximum suppression algorithm to complete the deduplication task of the detection results,and the objective function related to the detection accuracy and recall is also designed.It overcomes the greedy selection problem of the non-maximum suppression algorithm for high score detection results,and improves the detection ability of the algorithm for occluded environment.The performance of the proposed algorithm is verified on two datasets of PASCAL VOC and MS COCO,and the results show that the proposed algorithm effectively improves the average accuracy and location performance of the object detection task.Aiming at the problem that the local spatial information of 3D point cloud can not be effectively used,a point cloud encoding method based on local area is proposed.The axisaligned cube is used to search the local space area of the point cloud,and some points from the space area are selected to construct the encoding features of the points.A deep learning network is designed to deal with these encoding features of the point cloud.The new encoding feature not only contains abundant local spatial information of the point cloud,but also facilitates the deep learning network to extract the global spatial information of the point cloud,which improves the recognition ability of the network to 3D point cloud.The performance of the proposed method is verified on the Model Net40,S3 DIS,and Shape Net part datasets.The results show that the proposed algorithm effectively improves the recognition accuracy of deep learning networks in object classification,semantic segmentation,and part segmentation tasks.Aiming at the single object scene of the robotic grasp detection,a single-stage grasp detection algorithm based on region proposal networks is proposed.A five-dimensional grasp rectangular box is used to express the object grasp position,and an anchor matching strategy and a network loss function corresponding to the rectangular box are designed.The hard examples are used to train the proposed network,which overcomes the imbalance problem between positive and negative samples during training.The performance of the proposed algorithm is verified on Cornell and Jacquard grasp datasets.The results show that the proposed algorithm can effectively improve the grasp detection accuracy of the robot in the single object scene.Aiming at the robotic grasp detection scene with multiple objects stacked randomly,a multi-object grasp detection network is proposed based on object category detection network and single object grasp detection network.A matching strategy is designed to judge the dependency relationship between category detection and grasp detection results.The object detection network is used to detect the object category,and the single object grasp detection network is used to detect the graspable position of the objects,which decouples the two related detection tasks.The dense feature pyramid network is introduced to improve the ability of the detection network to utilize the low level location information.The performance of the proposed algorithm is verified on the VMRD multi-object grasp dataset.The results show that the proposed algorithm effectively improves the accuracy of the class detection and the accuracy of the grasp detection for different objects in cluttered scenes.Finally,the paper summarizes the whole work and innovation points,and points out the future research direction.