In the field of point cloud processing, deep learning is a mainstream method, but the existing methods do not fully utilize the local structure information of 3D point clouds, and have less local shape perception. We proposes a 3D point cloud processing model based on improved PoinetNet. Network model introduces position adaptive convolution into PointNet. The position?adaptive convolution constructs the kernel function by combining the weight matrices in the weight bank in a dynamic way, in which the coefficients of the weight matrix are adaptively learned from the relative positions of the points through the position?relative coefficient network. The kernel function constructed in this way can better solve the problem of irregularity and disorder of point cloud data. The classification accuracy of the position?adaptive network in the 3D object classification experiment is 3.60% higher than that of PointNet, and the average intersection ratio in the 3D object part segmentation experiment is 2.20% higher than that of PointNet. In the 3D scene semantics In the segmentation experiment, the average intersection and union ratio is improved by 9.14% compared with PointNet.

Taking advantage of its high?speed and high?precision characteristics, cooperative robots can improve production efficiency by imitating human creative and complex actions. At present, the simulation of human action mainly comes from the long?term debugging of the deployment personnel, which is lack of general solutions and can′t be deployed quickly. Based on this, an anchor?free RepVGG network?based Siamese network collaborative robot target tracking algorithm was proposed. The algorithm consists of a siamese network module, a classification regression module and a robot execution module. The siamese network module used the improved RepVGG network instead of ResNet as the backbone network to extract image features, which can improve the running speed of the whole network without losing accuracy，reduce the hardware requirements and is more friendly to special deep learning chips; the classification and regression module introduced the centrality branch to improve the prediction accuracy of the center point of the tracking frame; the robot execution module uses scale penalty and aspect ratio penalty to smooth the tracking boxes and ensure the smooth operation of the collaborative robots. Experimental results show that the average rate is 14 FPS higher than that before replacing the backbone network, which realizes real?time object tracking.

For most defogging algorithms, the effect on synthetic foggy images is different with the effect on real foggy images. Focusing on this problem, we propose a domain?adaptive single?image defogging algorithm which is based on a bidirectional conversion network. The bidirectional conversion network is designed to convert the foggy im?ages between the two domains. Our algorithm can be divided into two steps. Firstly, we reduce the difference be?tween the synthetic domain and the real domain by using the bidirectional conversion network; secondly, we remove fog from the input foggy images by using a convolutional neural network. To improve the effect and generalization ability of our algorithm, we use the synthetic RESIDE dataset and real foggy images as training data. Compared with results of some existing algorithms, the results show that our algorithm has better effects on foggy images of both domains, and our algorithm also improves the peak signal?to?noise ratio (PSNR) and structural similarity (SSIM).