Analysis of Rolling Shutter Correction Techniques for Visual SLAM

• Gabriel Pontarolo UFPR https://orcid.org/0009-0001-6416-4333
• Felipe Bombardelli UFPR https://orcid.org/0009-0004-5762-0849
• Eduardo Todt UFPR https://orcid.org/0000-0001-6045-1274

Resumo

Visual SLAM methods have become increasingly more popular in the latest years, largely due to their ability to achieve high accuracy while primarily relying on passive camera sensors. However, many of those algorithms restrict their scope to Global Shutter cameras, as the distortion caused by the Rolling Shutter readout can significantly degrade their performance. Still, CMOS cameras have become well established within the consumer market since they have lower cost and lower energy consumption compared to CCD. Recently, Rolling Shutter correction methods, notably those based on deep learning, have started to achieve remarkable accuracy. This work analyzes the performance of a visual odometry solution on multiple sequences of rolling shutter corrected images from the TUM RS-VIO dataset. Three state-of-the-art deep learning based correction methods were selected for analysis: 2F-DFRSC, JAMNet and SUNet. Results indicate that the sequences corrected by the first two algorithms are able to achieve a trajectory error comparable to those from Global Shutter sequence. Nevertheless, there is still room for optimization in terms of inference time before these solutions can be effectively used for real-time visual SLAM.

Palavras-chave: Deep learning, Visualization, Simultaneous localization and mapping, Cameras, Inference algorithms, Real-time systems, Trajectory, Sensors, Optimization, Visual odometry, Rolling Shutter, Visual SLAM, Computer Vision