LiDAR-Based Place Recognition in Dense Forests
Researchers analyzed the capabilities of four different LiDAR place recognition systems, both handcrafted and learning-based methods, using LiDAR data collected with a handheld device and legged robot within dense forest environments. Watch this video to see what they found.
Transcript
00:00:01 hello my name is hadam I will present evaluation and deployment of liab Base Place recognition in dance Forest liab Bas Place recognition methods in natural environment have been studied less closely compared to the urban environment like of distinctive structure loose foilage occlusions and limited field of view make natural environment particularly challenging in
00:00:23 this work we present evaluation and deployment of robust liab based Place recognition in dense forest environments we collected four different Forest data set for evaluation Evo and stum Ry are conifers wuts and for of or theous we use two different ligher models AI xc32 with 50 m effective range and 30° narrow field of fuel and a hassi QT 64 with 30 m effective range and 100°
00:00:54 wide fi of view and a frontier mapping payload mounted on a backpack our place recognition pipeline has two main modules State estimation and place recognition and verification server set estimation module provides odometry estimates at 10 HZ using liar and inara measurements initial Loop candidates are proposed for the most similar candidate in database We additionally verify
00:01:25 proposed Loop candidate using descripted distance and odometry course registration between Loop candidate pairs is done by ransac feature matching with local feature level verification checks finally IP find register Loop candidate pairs based on the proportion of inliers in MC the system either reject or accept the loop candidate if accepted we publish as a
00:01:50 loop closure constraint and optimize a post grab slam next we tested the performance of four different state of the art PL recognition methods look through a LD and scan context from Precision Reco curves look through theet out perform other models across four different Forest without experiencing any drastic drops in Precision even in the challenging scenario such as W mod
00:02:16 therefore we integrated the best performing method L 3D net to our system for further evaluation in online slam offline multia map merging and relocalization in this experiment we show place recognition on the all line slam operation the map is built incrementally as sensor moves inside the forest note that there are many frequent Loop closures inside of
00:02:42 forest also we obtain long Baseline distance up to 15 M between Loop closure pairs tree trunks are well defined with no evidence of drift Wii Ms are very dense decidious for trees are clotted with mixed pieces and ground vegetation as well as uneven terrain this is a challenging environment for light up Bas Place recognition due to a large proportion of
00:03:09 false Loop candidates our place recognition and verification server was able to detect and verify Loop candidates effectively rejecting false positives as a result we obtain consistent slam map minimizing the drift another use case is for multi Mission map merging multi- session map merging combines multiple sing Mission Maps into one unified map the goal is to
00:03:36 identify intermission Loop candidates in parly overlapping areas between each missions intermission Loop candidates are obtained through successive onetoone matching of individual mission to other Mission also intermission Loop candidates are detected within the forest without retracing exact path or relying on Open Access Road the resulting merged map allows us to
00:04:07 fit the terrain maps and extract the shape of individual trees by reconstruction the output tree models will be useful in relocalization scenario providing a prior map information about individual trees we showcase relocalization capability inside the forest and continuously track the pose of the lier sensor
00:04:30 the prior map was built using a back glider sequence a Forester is continuously localized within the forest while inspecting the trees this enables a Forester to visualize a rendering of the virtual tree model with important information while inspecting the trees in real time elect robot surveys the previously
00:04:53 mapped Forest by Forester a robot is continuously localized while inspecting the forest this application demonstrate the potential localization system to support Forest inventory or fure reality in summary we showcases Place recognition and verification server for robust Loop closure detection in online slam offline multiion map merging and
00:05:20 relocalization in online slam long Baseline Loop closures were captured inside the foret without following predefined path or relying on an Open Access Road offline multimission map merging enables us to find intermission Loop closures for large scale for mapping finally demonstration of the real localization application to the inspection task both performed by a
00:05:46 Forester and El robot thank you