Research On Object Grasping Detection Based On Convolutional Neural Network

Robotic grasp detection is an important technology that can help robotic systems perform efficient object grasping.Modern robot grasp detection technology can achieve fast and accurate grasping of complex objects by using deep learning algorithms and computer vision techniques.Although robot grasp detection techniques have made great progress,there are still some challenges,such as grasping complex objects,grasping dynamic objects,and grasping deformed objects.In addition,several factors need to be considered,such as the size,shape,weight,and surface texture of the object.While the robotic system also faces the influence of complex environmental factors such as noise,lighting,shadows,and reflections.Therefore,to address the problem of robustness of grasping detection algorithms in complex environments,especially in noisy environments,this thesis conducts the research on object grasping detection based on convolutional neural network algorithm,and the main research contents are as follows.(1)A convolutional neural network algorithm based on attentional feature fusion is proposed to cope with the impact of image noise on the object grasping and detection task.The algorithm is based on an encoding-decoding structure for end-to-end detection,and introduces an attention mechanism and a feature fusion module to enhance the extraction capability of image features for the effect of Gaussian noise on object images.The accuracy of the algorithm is verified by conducting experiments on publicly available datasets with an accuracy rate of 96.6%;meanwhile,the ablation experiments under the influence of different Gaussian noises demonstrate the robust performance and generalization ability of the algorithm.(2)A fused convolutional neural network algorithm combining wavelet variation and attention mechanism is proposed,aiming to solve the challenge of grasping detection in a multi-noise environment.The algorithm improves the U-Net network structure by fusing wavelet transform and convolutional neural network to solve the noise interference by using the noise reduction property of wavelet transform.The algorithm continues the application of attention mechanism,which makes the feature extraction ability of the network enhanced.The algorithm uses depth separable convolution instead of traditional convolution to significantly reduce the parameter scale of the network.In addition,the diversity and randomness of the noise are extended.Through experiments,it is shown that the accuracy of the algorithm is improved to 97.75%,and the accuracy of grasping detection is further improved;for the effect of multiple noises,the excellent performance is still maintained,and the robustness is further enhanced.(3)The robot grasping experiments in a real environment were carried out.Firstly,the robot grasping experiment platform was built,and the overall construction of the grasping platform and the composition of the hardware and software system were described;then the vision system of the grasping platform was calibrated;then the robustness of the algorithm was experimentally explored;finally,the grasping experiment was conducted for unknown objects,and the algorithm was evaluated in terms of the recognition of the generated grasping frame and the real grasping of the objects,divided into two scenarios: single object and multi-object clutter.The effectiveness of the algorithm is evaluated in two scenarios: single-object and multi-object clutter.The results of the grasping experiments show that the proposed algorithm is effective and has high robustness performance in real environment grasping.