Hard-Lite SLAM: A Hybrid Detector Based Real-Time SLAM System

Abstract

Simultaneous Localization and Mapping (SLAM) system is essential for autonomous driving and mobile robots. The problem of data association between features becomes a bottleneck limiting the performance of traditional visual SLAM systems, especially in complex environments. Therefore, many studies combine the SLAM system with the Convolutional Neural Networks (CNN) to obtain a more robust data association. This paper shows that CNN-based local descriptors significantly improve the accuracy and robustness of the SLAM system. The CNN-based keypoints reduce the performance of the SLAM algorithm in many scenarios. We propose a SLAM system that combines hand-crafted keypoints with CNN local descriptors. The system is more robust in complex environments than traditional visual SLAM systems. The experimental results show that our system achieves higher localization accuracy than ORB-SLAM2 and VINS-Mono on the evaluated datasets. Meanwhile, the CNN local descriptors can be combined with any visual SLAM system and have good portability. Furthermore, with the assistant of the Nvidia TensorRT inference acceleration technology, the system can run in real-time on the Jetson AGX Xavier at 27 frames per second. CCS CONCEPTS • Computer systems organization • Embedded and cyber-physical systems • Robotics • Robotic autonomy