I’m still evaluating the performance of the global localization. I recorded a ROS bag that includes the VSLAM path, the camera topic, and the global localization results.
28-04-2025-path-rosbag-zimmer-2.zip (61.0 MB)
While playing the rosbag, I observed via the QRT console logger that some poses were successfully identified. However, the global pose (/visual_localization/pose) appears to jump around inconsistently. Below are the launch file settings I used:
‘map_dir’: map_directory,
‘config_dir’: config_directory,
‘num_cameras’: 2,
‘publish_map_to_base_tf’: True,
‘verbose_logging’: True,
‘glog_v’: 1,
‘init_glog’: True,
‘image_topics’: image_topic_list,
‘camera_info_topics’: camera_info_topic_list,
‘camera_names’: camera_names_value,
‘localization_precision_level’: 2,
‘map_frame’: “map”,
‘base_frame’: “base_link”,
‘debug_dir’: debug_directory,
‘enable_continuous_localization’: True
The global localization map contains 832 keyframes with a total of 1,069,881 features, covering an area of approximately 13 m². I extracted the keyframe positions from the frames_meta.json file generated during map creation to analyze their distribution.
For global localization, I’m using a single stereo camera. Based on the number of visual words and the density of keyframes, i think the map should be sufficient for reliable localization.
Does anyone have advice on how to improve the functionality of the global localization, or can anyone spot the mistake I might be making here?