Data-Efficient Decentralized Visual SLAM

Abstract

Decentralized visual simultaneous localization and mapping (SLAM) is a powerful tool for multi-robot applications in environments where absolute positioning systems are not available. Being visual, it relies on cameras, cheap, lightweight and versatile sensors, and being decentralized, it does not rely on communication to a central ground station. In this work, we integrate state-of-the-art decentralized SLAM components into a new, complete decentralized visual SLAM system. To allow for data association and co-optimization, existing decentralized visual SLAM systems regularly exchange the full map data between all robots, incurring large data transfers at a complexity that scales quadratically with the robot count. In contrast, our method performs efficient data association in two stages: in the first stage a compact full-image descriptor is deterministically sent to only one robot. In the second stage, which is only executed if the first stage succeeded, the data required for relative pose estimation is sent, again to only one robot. Thus, data association scales linearly with the robot count and uses highly compact place representations. For optimization, a state-of-the-art decentralized pose-graph optimization method is used. It exchanges a minimum amount of data which is linear with trajectory overlap. We characterize the resulting system and identify bottlenecks in its components. The system is evaluated on publically available data and we provide open access to the code.

DOOR-SLAM: Distributed, Online, and Outlier Resilient SLAM for Robotic Teams

“…In the frontend, we use the latest version of RTAB-Map [45] for stereo visual odometry and we use the tensorflow implementation of NetVLAD provided in [9], with the default neural network weights trained in the original paper [10]. We only keep the first 128 dimensions of the generated descriptors to limit the data to be exchanged, as done in [9]. The visual feature extraction and relative pose transformation estimation are done by adapting the implementation in RTAB-Map and keeping their default parameters.…”

“…We used a PCM threshold of 1%, a NetVLAD comparison threshold of 0.15, and a minimum of 5 feature correspondences in the geometric verification to get a high number of loop closure measurements. While related work uses more conservative thresholds for NetVLAD and the number of feature correspondences to avoid outliers [9], we can afford more aggressive thresholds thanks to PCM. Results. Fig.…”

“…The second module is a data-efficient distributed SLAM front-end. Similar to the recent approach [9], our system uses NetVLAD descriptors [10] for place recognition. However, our approach does not require full connectivity maintenance nor additional environment-specific pre-training.…”

Search and Rescue Under the Forest Canopy Using Multiple UAS

“…On the other hand, Dong et al (2015), Morrison et al (2016) and Schuster et al (2019) propose to run full SLAM onboard each vehicle. The incurred computation costs are further reduced in distributed architectures, where each robot only optimizes its local map and shares the compressed map or boundary poses with each other, see Cunningham et al (2013); Choudhary et al (2017); Cieslewski et al (2018).…”

Geometrically Mappable Image Features

“…To use the training script for images instead, we replace all point cloud specific loaders and containers with their image counterparts. We replace the network to be trained with the VGG-16 and NetVLAD architecture implementation of [26]. Furthermore, we use the training parameters specified in [19], reducing the learning rate to 0.000001, the number of queries per batch to two, and the number of positives and negatives per query to six each.…”