Non-Iterative SLAM (Best Paper Award in ICAR 2017)

Non-Iterative slam is a novel visual slam algorithm, that is very light and can be run on ultra low power CPU. It leverages on inertial sensors and any 3-D sensors that can provide 3-D point clouds such as stereo camera, depth camera and 3-D laser scanner On the Intel Real Sense Robotic Development Kit that features a mobile-level processor. Non-Iterative SLAM can still achieve high update frequency with high map resolution. Typically, higher resolution will result in slower update rate. However, Non-Iterative SLAM still achieves the faster update rate with the high map resolution using an ultra-low power CPU without GPU devices. It is designed for micro unmanned aerial vehicles that only provide limited payloads But it also can be used for other scenarios The reason we design Non-Iterative SLAM is that iterative solutions have become the bottleneck of traditional methods Nearly all the existing methods need iterative solutions which are very time consuming For example, they need RANSAC to remove outliers, need ICP to match point clouds. What’s more, feature-based methods and direct methods need iterative solutions, such as Gaussian-Newton method to minimize reprojection error or photometric error respectively. Unfortunately, they are sensitive to initialization and cannot guarantee the global optimum. The deep learning based methods also require iterative solutions to find the optimal parameters of the neural network and require large number of training data To overcome this problem, we propose a new framework for data association It based on single key-frame training and can be conducted online Therefore, it doesn’t need any prior knowledge Instead of using any iterative techniques, we find a closed-form solution for the new objective function Therefore, our proposed method can be run very fast and very robust to fast motion To prove this, we compare our algorithm with Google’s Projects Tango Google’s project Tango aims to provide accurate motion tracking. During the fast shaking motion, the number of tracking failure is counted. It should be noted that the tango tablet also depends on built-in GPU device to do feature matching But we only need a low-power CPU These experiments show that Non-Iterative SLAM can track very fast motion very well But Google’s Project Tango always fails in this situation

One Reply to “Non-Iterative SLAM (Best Paper Award in ICAR 2017)”

Leave a Reply

Your email address will not be published. Required fields are marked *