DFE-SLAM: Dynamic SLAM based on improved feature extraction

Traditional visual localization algorithms often assume a static world, making them susceptible to inaccuracies and reduced robustness in real environments with dynamic objects. Additionally, these algorithms struggle to maintain high localization accuracy in low-texture settings. To overcome these challenges, we propose a dynamic SLAM system with enhanced feature extraction capabilities, called DFE-SLAM (Dynamic SLAM based on improved feature extraction). DFE-SLAM employs a lightweight YOLOv5s dynamic target detection network to filter out dynamic features from detection results. It also utilizes adaptive thresholding during the feature point detection stage to enhance accuracy. Moreover, the system incorporates a dense map construction thread, which creates dense point cloud maps using keyframes with corresponding bit poses. We evaluated DFE-SLAM using the TUM and Bonn datasets, benchmarking it against several advanced SLAM systems. The results demonstrate that DFE-SLAM excels in dynamic environments, delivering superior accuracy and robustness compared to most existing SLAM systems.