Leveraging 2D Annotations for Cost-Effective Dynamic Urban Scene Reconstruction

With the development of enterprise intelligence and automation, the inspection of power distribution room is not limited to manual work. This paper proposes a semantic segmentation method based on improved ORB-SLAM3 and YOLOV5 fusion to eliminate dynamic features. ORB-SLAM3 is used to realize scene reconstruction to generate dense point cloud maps. In this paper, all the differences between ORB-SLAM2 and ORBSLAM3 are compared, and the dense point cloud generation and loop detection modules are added to make the mapping more accurate. In real scenes, the depth sensor is inevitably disturbed by dynamic objects when obtaining real environmental information. Compared with dynamic SLAM for object detection, the semantic segmentation module of YOLOV5 is used in this paper, which can remove dynamic objects while minimizing the loss of effective features. In addition, TensorRT is used to accelerate the construction of the map, and the semantic map is generated, which makes a prerequisite for the subsequent conversion of the map into a two-dimensional occupancy map. Experimental results show that compared with the original ORB-SLAM2/3, the improved ORBSLAM3 greatly improves the accuracy in dynamic scenes. A more real-time and accurate scene reconstruction function of power distribution room is realized.

Object tracking and 3D reconstruction are often performed together, with tracking used as input for reconstruction. However, the obtained reconstructions also provide useful information for improving tracking. We propose a novel method that closes this loop, first tracking to reconstruct, and then reconstructing to track. Our approach, MOTSFusion (Multi-Object Tracking, Segmentation and dynamic object Fusion), exploits the 3D motion extracted from dynamic object reconstructions to track objects through long periods of complete occlusion and to recover missing detections. Our approach first builds up short tracklets using 2D optical flow, and then fuses these into dynamic 3D object reconstructions. The precise 3D object motion of these reconstructions is used to merge tracklets through occlusion into long-term tracks, and to locate objects when detections are missing. On KITTI, our reconstruction-based tracking reduces the number of ID switches of the initial tracklets by more than 50%, and outperforms all previous approaches for both bounding box and segmentation tracking. Code available: <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/tobiasfshr/MOTSFusion</uri> .

Monocular depth reconstruction of complex and dynamic scenes is a highly challenging problem. While for rigid scenes learning-based methods have been offering promising results even in unsupervised cases, there exists little to no literature addressing the same for dynamic and deformable scenes. In this work, we present an unsupervised monocular framework for dense depth estimation of dynamic scenes, which jointly reconstructs rigid and nonrigid parts without explicitly modelling the camera motion. Using dense correspondences, we derive a training objective that aims to opportunistically preserve pairwise distances between reconstructed 3D points. In this process, the dense depth map is learned implicitly using the as-rigid-as-possible hypothesis. Our method provides promising results, demonstrating its capability of reconstructing 3D from challenging videos of non-rigid scenes. Furthermore, the proposed method also provides unsupervised motion segmentation results as an auxiliary output.

LiDAR point cloud is essential for autonomous vehicles, but motion distortions from dynamic objects degrade the data quality. While previous work has considered distortions caused by ego motion, distortions caused by other moving objects remain largely overlooked, leading to errors in object shape and position. This distortion is particularly pronounced in high-speed environments such as highways and in multi-LiDAR configurations, a common setup for heavy vehicles. To address this challenge, we introduce HiMo, a pipeline that repurposes scene flow estimation for non-ego motion compensation, correcting the representation of dynamic objects in point clouds. During the development of HiMo, we observed that existing self-supervised scene flow estimators often produce degenerate or inconsistent estimates under high-speed distortion. We further propose SeFlow++, a real-time scene flow estimator that achieves state-of-the-art performance on both scene flow and motion compensation. Since well-established motion distortion metrics are absent in the literature, we introduce two evaluation metrics: compensation accuracy at a point level and shape similarity of objects. We validate HiMo through extensive experiments on Argoverse 2, ZOD and a newly collected real-world dataset featuring highway driving and multi-LiDAR-equipped heavy vehicles. Our findings show that HiMo improves the geometric consistency and visual fidelity of dynamic objects in LiDAR point clouds, benefiting downstream tasks such as semantic segmentation and 3D detection. See https://kin-zhang.github.io/HiMo for more details.

We propose a content-based 3D mosaic representation for long video sequences of 3D and dynamic scenes captured by a camera on a mobile platform. The motion of the camera has a dominant direction of motion (as on an airplane or ground vehicle), but 6 degrees-of-freedom (DOF) motion is allowed. In the first step, a pair of generalized parallel-perspective (pushbroom) stereo mosaics is generated that captured both the 3D and dynamic aspects of the scene under the camera coverage. In the second step, a segmentation-based stereo matching algorithm is applied to extract parametric representation of the color, structure and motion of the dynamic and/or 3D objects in urban scenes where a lot of planar surfaces exist. Based on these results, the content-based 3D mosaic (CB3M) representation is created, which is a highly compressed visual representation for very long video sequences of dynamic 3D scenes. Experimental results are given

Chapter 2 - Multi-View Calibration, Synchronization, and Dynamic Scene Reconstruction

Visual simultaneous localization and mapping (SLAM) in dynamic environments has received significant attention in recent years, and accurate segmentation of real dynamic objects is the key to enhancing the accuracy of pose estimation in such environments. In this study, we propose a visual SLAM approach based on ORB-SLAM3, namely dynamic multiple object tracking SLAM (DMOT-SLAM), which can accurately estimate the camera’s pose in dynamic environments while tracking the trajectories of moving objects. We introduce a spatial point correlation constraint and combine it with instance segmentation and epipolar constraint to identify dynamic objects. We integrate the proposed motion check method into DeepSort, an object tracking algorithm, to facilitate inter-frame tracking of dynamic objects. This integration not only enhances the stability of dynamic features detection but also enables the estimation of global motion trajectories for dynamic objects and the construction of object-level semi-dense semantic maps. We evaluate our approach on the public TUM, Bonn, and KITTI dataset, and the results show that our approach has a significant improvement over ORB-SLAM3 in dynamic scenes and performs better compared to other state-of-the-art SLAM approaches. Moreover, experiments in real-world scenarios further substantiate the effectiveness of our approach.

Visual simultaneous localization and mapping (VSLAM) is extensively utilized in the field of unmanned driving. Classical VSLAM typically postulates that the surroundings are stationary, and dynamic objects can seriously affect the accuracy of the system. However, detecting and tracking dynamic objects is beneficial for improving the autonomy and intelligence of robotics. In order to solve the problem of VSLAM distinguishing dynamic objects in dynamic scenes, a semantic VSLAM system based on the VDO-SLAM framework is proposed. The approach integrates semantic segmentation into the visual odometry and combines scene flow and geometric constraints to effectively reduce the negative impact of dynamic objects on system performance. Specifically, a transformer-based neural network architecture for optical flow estimation was employed, which effectively enhances the segmentation precision of foreground and background regions. And a geometric constraint was introduced to classify dynamic objects, which can reduction of erroneous segmentation of dynamic objects with effect. The data analysis results on the KITTI dataset demonstrate that our proposed algorithms can effectively enhance precision and stability of VSLAM system, reduce the tracking error of VSLAM in dynamic environments, and demonstrate good potential and application value.

We propose a content-based 3D mosaic (CB3M) representation for long video sequences of 3D and dynamic scenes captured by a camera on a mobile platform. After a set of parallel-perspective (pushbroom) mosaics with varying viewing directions is generated, a multi-view, segmentation-based stereo matching algorithm is applied to extract parametric representations of the color, structure and motion of the dynamic and/or 3D objects in urban scenes. We use the fact that all the static objects obey the epipolar geometry of pushbroom stereo, whereas an independent moving object either violates the epipolar geometry if the motion is not in the direction of sensor motion or exhibits unusual 3D structures. The CB3M is a highly compressed visual representation for a very long video sequence of a dynamic 3D scene. More importantly, the CB3M representation has object contents of both 3D and motion. Experimental results are given for the CB3M construction for both simulated and real video sequences.

We present a method for dynamic surface reconstruction of large-scale urban scenes from LiDAR. Depth-based re-constructions tend to focus on small-scale objects or large-scale SLAM reconstructions that treat moving objects as outliers. We take a holistic perspective and optimize a compositional model of a dynamic scene that decomposes the world into rigidly-moving objects and the background. To achieve this, we take inspiration from recent novel view syn-thesis methods and frame the reconstruction problem as a global optimization over neural surfaces, ego poses, and object poses, which minimizes the error between composed spacetime surfaces and input LiDAR scans. In contrast to view synthesis methods, which typically minimize 2D errors with gradient descent, we minimize a 3D point-to-surface error by coordinate descent, which we decompose into reg-istration and surface reconstruction steps. Each step can be handled well by off-the-shelf methods without any re-training. We analyze the surface reconstruction step for rolling-shutter LiDARs, and show that deskewing operations common in continuous time SLAM can be applied to dynamic objects as well, improving results over prior art by an order of magnitude. Beyond pursuing dynamic reconstruction as a goal in and of itself, we propose that such a system can be used to auto-label partially annotated se-quences and produce ground truth annotation for hard-to-label problems such as depth completion and scene flow. Please refer https://anishmadan23.github.io/smore/ for more visual results.

Urban areas are rapidly changing all over the world and therefore provoke the necessity to update urban maps frequently. Remote sensing has been used for many years to monitor these changes. The urban scene is characterized by a very high complexity, containing objects formed from different types of man-made materials as well as natural vegetation. Hyperspectral sensors provide the capability to map the surface materials present in the scene using their spectra and therefore to identify the main object classes in the scene in a relatively easy manner. However ambiguities persist where different types of objects are constructed of the same material. This is for instance the case for roads and roof covers. Although higher-level image processing (e.g. spatial reasoning) might be able to relief some of these constraints, this task is far from straight forward. In the current paper the authors fused information gathered using a hyperspectral sensor with that of high-resolution polarimetric SAR data. SAR data give information about the type of scattering backscatter from an object in the scene, its geometry and its dielectric properties. Therefore, the information obtained using the SAR processing is complementary to that obtained using hyperspectral data. This research was applied on a dataset consisting of hyperspectral data from the HyMAP sensor (126 channels in VIS-SWIR) and E-SAR data which consists of fullpolarimetric L-band and dual-polarisation (HH and VV) X-band data. Two supervised classifications are used; 'Logistic Regression' (LR) which applied to the SAR and the PolSAR data and a 'Matched Filter' which is applied to the hyperspectral data. The results of the classification are fused in order to improve the mapping of the main classes in the scene and were compared to a ground truth map that was constructed by combining a digital topographic map and a vectorized cadastral map of the research area. An adequate change detection of man-made objects in urban scenes was obtained by the fusion of features derived from SAR, PolSAR and hyperspectral data.

While dynamic scene reconstruction has made revolutionary progress from the earliest setup using a mass of static cameras in studio environment to the latest egocentric or hand-held moving camera based schemes, it is still restricted by the recording volume, user comfortability, human labor and expertise. In this paper, a novel solution is proposed through a real-time and robust dynamic fusion scheme using a single flying depth camera, denoted as FlyFusion. By proposing a novel topology compactness strategy for effectively regularizing the complex topology changes, and the Geometry And Motion Energy (GAME) metric for guiding the viewpoint optimization in the volumetric space, FlyFusion succeeds to enable intelligent viewpoint selection based on the immediate dynamic reconstruction result. The merit of FlyFusion lies in its concurrent robustness, efficiency, and adaptation in producing fused and denoised 3D geometry and motions of a moving target interacting with different non-rigid objects in a large space.

We demonstrate the use of shape-from-shading (SfS) to improve both the quality and the robustness of 3D reconstruction of dynamic objects captured by a single camera. Unlike previous approaches that made use of SfS as a post-processing step, we offer a principled integrated approach that solves dynamic object tracking and reconstruction and SfS as a single unified cost function. Moving beyond Lambertian S f S , we propose a general approach that models both specularities and shading while simultaneously tracking and reconstructing general dynamic objects. Solving these problems jointly prevents the kinds of tracking failures which can not be recovered from by pipeline approaches. We show state-of-the-art results both qualitatively and quantitatively.

Simultaneous localization and map building (SLAM) is crucial in autonomous robot navigation. However, existing SLAM systems generally assume a static environment, which makes it difficult to cope with the interference caused by moving objects in dynamic scenes, affecting the system’s localization accuracy and robustness. To address this challenge, this paper proposes YKD-SLAM, a visual SLAM system for indoor dynamic environments, which is based on the ORB-SLAM2 framework and incorporates YOLOv8 target detection, RCF-KMeans (Region-ConstrainedFastK-Means), and epipolar geometric constraints to realize the accurate rejection of dynamic feature points and improve the localization performance in dynamic environments. YKD-SLAM first uses YOLOv8 to detect dynamic objects in the scene, generates a detection frame, optimizes the depth map through open operations, and performs multi-region segmentation of the region within the detection frame by combining RCF-KMeans. Subsequently, through the dynamic feature point rejection strategy based on epipolar geometric constraints, different regions in the detection frame are discriminated into dynamic and static regions, and the feature points in the dynamic region are rejected to improve the localization accuracy and robustness of the system in dynamic environments. The experimental results show that YKD-SLAM performs well in several dynamic scenes in the TUMRGB-D dataset. Compared with ORB-SLAM2, its ATE is reduced by 98.37%; compared with DynaSLAM, the system operation efficiency is improved by 95.35%. In addition, practical experiments conducted in indoor dynamic scenes further validate its potential in real applications.

This paper addresses the challenge that current event-based video reconstruction methods cannot produce static background information. Recent research has uncovered the potential of event cameras in capturing static scenes. Nonetheless, image quality deteriorates due to noise interference and detail loss, failing to provide reliable background information. We propose a two-stage reconstruction strategy to address these challenges and reconstruct static scene images comparable to frame cameras. Building on this, we introduce the URSEE framework designed for reconstructing motion videos with static backgrounds. This framework includes a parallel channel that can simultaneously process static and dynamic events, and a network module designed to reconstruct videos encompassing both static and dynamic scenes in an end-to-end manner. We also collect a real-captured dataset for static reconstruction, containing both indoor and outdoor scenes. Comparison results indicate that the proposed method achieves state-of-the-art performance on both synthetic and real data.