PIDNET-SLAM: A MULTI-RESOLUTION SEMANTIC SLAM ALGORITHM FOR DYNAMIC SCENES

Abstract

Simultaneous Localization and Mapping (SLAM) enables robots to map their surroundings while determining their positions in real time. Traditional SLAM systems typically assume static environments, causing accuracy issues when dynamic objects are present. Existing methods, which often use deep learning techniques like semantic segmentation, instance segmentation, and object detection, increase accuracy but add significant computational overhead, making them unsuitable for real time use. Therefore, a robust and computationally efficient SLAM solution capable of effectively handling dynamic scenes is needed.

To address these challenges, we introduce PIDNet SLAM, an innovative extension of ORB SLAM 3 that effectively manages dynamic environments. PIDNet SLAM incorporates a multi resolution semantic segmentation network based on PIDNet, consisting of two parallel branches: a low resolution branch actively processing non keyframes, and a high resolution branch providing supplementary high resolution features without interfering with the main workflow. Additionally, a lightweight geometric module uses geometric transformations and optical flow to identify and eliminate dynamic features, enhancing the system’s overall accuracy.

Evaluations on the dynamic TUMM RGB-D dataset show PIDNet SLAM significantly outperforming existing methods like Dyna SLAM, SOLO SLAM, and ORB SLAM 3, achieving over 97% improvement in localization accuracy in dynamic scenarios. Remarkably, the system maintains an average processing time of only 49ms per frame on a low power GPU, demonstrating an optimal balance between accuracy and computational efficiency. PIDNet SLAM thus presents a critical advancement toward effective real time SLAM in dynamic environments, with potential future improvements focusing on refining the low resolution branch for even faster performance.