A lightweight model of pose estimation based on DenseFusion

6D pose estimation is a key research problem in the field of artificial intelligence, with many applications in robot grasping, autonomous navigation, and reality augmentation. Due to the specificity of the deployment environment, it is often difficult to apply bit-pose estimation models with a large number of parameters to resource-constrained mobile devices such as smartphones and wearable smart devices. The common pose estimation network DenseFusion is based on the collaborative processing of RGB data and depth data, which can efficiently fuse the extracted texture features and depth features together, effectively improving the accuracy of 6D pose estimation. To solve the problem of oversized parameters of the bit-pose estimation model, the basic principle of the deep neural network model is used to lightly compress the model based on the existing compression algorithm, which uses a fusion of structured pruning strategy and unstructured pruning iterations according to the characteristics of the model network. The experimental data, tested and validated on the Linemod dataset and the home-made dataset, show that the storage capacity of the model compression decreases to 70.7MB using the improved method of this experiment, and the accuracy rate is maintained at around 90%. This shows that the lightweight DenseFusion model has a significantly lower number of model parameters, and its accuracy fluctuates little, reducing the requirement for hardware platforms.