Efficient structural perception is essential for mapping and autonomous navigation on resource-constrained robots. Existing 3D methods are computationally prohibitive, while traditional 2D geometric approaches lack robustness. This paper presents a lightweight, real-time framework that projects 3D LiDAR data into 2D Bird's-Eye-View (BEV) images to enable efficient detection of structural elements relevant to mapping and navigation. Within this representation, we systematically evaluate several feature extraction strategies, including classical geometric techniques (Hough Transform, RANSAC, and LSD) and a deep learning detector based on YOLO-OBB. The resulting detections are integrated through a spatiotemporal fusion module that improves stability and robustness across consecutive frames. Experiments conducted on a standard mobile robotic platform highlight clear performance trade-offs. Classical methods such as Hough and LSD provide fast responses but exhibit strong sensitivity to noise, with LSD producing excessive segment fragmentation that leads to system congestion. RANSAC offers improved robustness but fails to meet real-time constraints. In contrast, the YOLO-OBB-based approach achieves the best balance between robustness and computational efficiency, maintaining an end-to-end latency (satisfying 10 Hz operation) while effectively filtering cluttered observations in a low-power single-board computer (SBC) without using GPU acceleration. The main contribution of this work is a computationally efficient BEV-based perception pipeline enabling reliable real-time structural detection from 3D LiDAR on resource-constrained robotic platforms that cannot rely on GPU-intensive processing. The source code and pre-trained models are publicly available.
Figures & tables
Figure 1: Wall and Room Recognition Module Architecture. The pipeline details the modular process of transforming raw 3D LiDAR data into semantic structural elements via 2D projection and geometric feature extraction.
Figure 2: System Architecture Diagram. The system comprises five ROS2 nodes, categorized into three functional modules based on the data processing pipeline: Data Interface, Feature Detection, and Feature Fusion.
Figure 3: Qualitative Comparison between a Real LiDAR BEV Scan and a Generated Synthetic Training Sample. The correspondence arrows illustrate the high fidelity of the proposed generation pipeline in mimicking real-world sensor characteristics: (Top) Continuous wall structures are automatically annotated with green OBBs (Class 0); (Middle) Real-world noise patterns are injected to simulate sensor artifacts; (Bottom) Discrete obstacles like columns are instantiated with red OBBs (Class 1). This visual similarity ensures a minimal domain gap for model training.
Figure 4: Flowchart of the Data Post-processing and Fusion Algorithm. The process is divided into two main stages: the Local Fusion Node for feature extraction and spatial merging, and the Global Fusion Node for multi-epoch data fusion and noise filtering.
Figure 5: Merge Wall Envelopes for LSD and Hough Transform. Illustration of the merge wall envelopes step, consolidating double-edge structures into a single representative line.
Figure 6: Illustration of the CLustering Metrics. Depiction of the three components used in the custom distance metric for DBSCAN: spatial distance Δd , angular difference Δθ , and segment overlap Δo .
Figure 7: Pipeline of the Manhattan World Optimizer. Leveraging the inherent Manhattan structural prior of indoor scenes, this optimizer integrates scattered semantic information of wall segments into a high-precision, topologically closed indoor room floorplan.
Figure 8: Figures of four scenarios. The pipeline is tested within the four scenarios, which range from a few meters to tens of meters.
Figure 9: Performance Comparison of the Four Methods in Garage Scenario across Three Processing Stage. The figure shows the outputs of three nodes: Detector, Local Fusion, and Global Fusion. For Detector Output, the black 2D point cloud represents the 3D LiDAR BEV map, overlaid with wall recognition results. Notably, for RANSAC, green marking indicate line segments rejected as too short, while for YOLO, cyan markings identify columns. For Local Fusion Output, blue line segments represent the raw input received from the Detector Node, while red segments depict the spatially fused wall structures. For Global Fusion Output, the blue segments correspond to the input from the Local Fusion Node, and red segments illustrate the final temporally fused global map. Numerical annotations on the segments indicate the rank of the observation matrix, reflecting estimation stability.
Figure 10: Global Fusion Outputs for the Corridor Scenario. The presence of glass surfaces causes specular reflections, resulting in geometric artifacts resembling a parallel corridor.
Figure 11: Raw Point Cloud and Global Fusion Outputs for the Laboratory Scenario. The left panel displays the raw 3D point cloud. The right panels display a 2 × 2 comparison of the four methods against the ground truth. Legend: Green lines represent the Ground Truth; Pink lines indicate successfully matched detections (True Positives); Blue lines represent unmatched detections (False Positives); Orange lines correspond to the matched segments of the Ground Truth.
Figure 12: Global Fusion Outputs for the Classroom Hallway Scenario. The comparison of outputs against the ground truth is shown as well. While geometric methods capture fine details, the deep learning approach exhibits lower recall in this specific setting.
Figure 13: Latency Breakdown Per Frame in the Garage Scenario. The stacked bars illustrate the time consumption of the Detector, Local Fusion, Global Fusion, and Data Transfer components. The corresponding methods include the Hough Transform, LSD, RANSAC, and the YOLO-OBB method with image size of 1024 and 640. The red dashed line represents the 10 Hz real-time threshold (100 ms).
Figure 14: Average resource consumption comparison on the mobile computing platform. (Left) Average CPU usage percentage, where RANSAC exhibits excessive load due to iterative calculations. (Right) System memory usage percentage. Although YOLO-OBB requires higher memory allocation for the neural network, they remain below 7% of the total available RAM (8GB), proving its feasibility for onboard deployment.
Figure 15: Detection Performance Evaluation across Four Scenarios. (Left) Length-Weighted Recall scores. (Right) Length-Weighted Precision scores. Error bars indicate the standard deviation. YOLO-OBB demonstrates superior robustness, particularly in the cluttered Laboratory scenario, while maintaining high precision across all environments.
Method
Scenario
Garage
Corridor
Laboratory
Classroom Hallway
Hough Transform
0.86 ± 0.09
0.73 ± 0.08
0.61 ± 0.23
0.53 ± 0.08
LSD
0.66 ± 0.07
0.52 ± 0.05
0.58 ± 0.12
0.43 ± 0.05
RANSAC
0.67 ± 0.03
0.61 ± 0.02
0.39 ± 0.00
0.18 ± 0.02
YOLO-OBB
0.85 ± 0.01
0.72 ± 0.01
0.84 ± 0.05
0.62 ± 0.04
Table 1: Quantitative Detection Performance (F1-Score) across Scenarios. The table presents the average and standrad deviation of F1-Score to evaluate the overall detection quality.
Scenario
Distance Error (cm) ↓
Angle Error (deg) ↓
Hough
LSD
RANSAC
YOLO
Hough
LSD
RANSAC
YOLO
Garage
6.47
4.65
3.47
6.03
0.25
0.23
0.14
0.09
Corridor
10.36
7.03
5.34
6.15
0.41
0.24
0.10
0.35
Laboratory
13.16
6.80
-
9.38
1.52
1.42
-
1.09
Classroom Hallway
10.17
15.38
-
5.73
0.95
0.73
-
1.12
Average
10.02
8.46
N/A
6.83
0.77
0.66
N/A
0.66
Table 2: Geometric Accuracy Assessment of the Reconstructed Maps. The table compares the Perpendicular Distance Error (cm) and Angular Error (degrees) across four scenarios. Note that “-” indicates an algorithm failure where insufficient line segments were matched (specifically RANSAC in complex environments).
Method
Large Simple Scenario
Complex Scenario
Robustness
Hough
Efficient; lacks semantic awareness.
High false positives due to clutter.
Medium
LSD
High geometric detail; fast.
Severe over-segmentation; computational bottleneck.
Medium
RANSAC
Good modeling; high latency.
Convergence failure due to low inlier ratios.
Low
YOLO
Accurate classification of structural elements.
Excellent clutter rejection ; high stability.
High
Table 3: Qualitative Summary of Method Performance
Figure 16: Over-segmentation of the LSD Method. (Left) It shows the original point cloud BEV map directly processed by the LSD method. (Right) It presents the corresponding extracted line segments, which are extremely numerous.
Figure 17: False Negative Detection due to Feature Resolution Mismatch. The figure illustrates a specific failure case in Classroom Hallway, a small-scale scenario. The blue lines represent the model’s predictions overlaid on the raw point cloud BEV. The red circle highlights a distinct, thin wall segment that the model failed to detect (False Negative). This omission is attributed to the domain gap between the thick, noisy walls in the synthetic training data and the sharp, thin features in this real-world scan.
Efficient indoor LiDAR perception is challenging because mobile robots must understand cluttered three-dimensional environments under strict latency and memory constraints. Existing point-based and voxel-based methods often incur substantial computational overhead, whereas conventional bird's-eye-view (BEV) representations improve efficiency at the cost of discarding vertical geometric information. We present IndoorBEV, a lightweight LiDAR perception framework that mitigates this tradeoff through a height-aware BEV representation and geometry-conditioned feature fusion. IndoorBEV summarizes the vertical point distribution in each BEV cell using statistical height features and multi-frequency height encoding, allowing informative three-dimensional cues to be processed efficiently by two-dimensional convolutions. A lightweight encoder then integrates complementary geometric features with multi-scale local representations and compact global scene context. Decoupled dense prediction heads jointly produce semantic BEV maps and oriented object bounding boxes. IndoorBEV contains only 0.6M parameters and requires 2.3 MB of model storage. On an NVIDIA AGX Orin, it uses 21.52 MB of GPU memory per inference and achieves a mean latency of 169.6 ms under a 200 ms perception deadline, with a deadline miss ratio of 1.8%. Evaluations on simulated scenes, real-world robot scans, and an open-source indoor point-cloud dataset demonstrate a favorable tradeoff among perception accuracy, latency, and memory consumption. These results indicate that explicitly encoding vertical geometry within a compact BEV representation provides an effective approach to resource-efficient indoor LiDAR perception.
Safe autonomous agents and mobile robots need fast real time 3D perception, especially for vulnerable road users (VRUs) such as pedestrians. We introduce a new bird's eye view (BEV) encoding, which maps the full 3D LiDAR point cloud into a light-weight 2D BEV tensor with three height bands. We explicitly reformulate 3D detection as a 2D detection problem and then reconstruct 3D boxes from the BEV outputs. A single network detects cars, pedestrians, and cyclists in one pass. The backbone uses area attention at deep stages, a hierarchical bidirectional neck over P1 to P4 fuses context and detail, and the head predicts oriented boxes with distribution focal learning for side offsets and a rotated IoU loss. Training applies a small vertical re bin and a mild reflectance jitter in channel space to resist memorization. We use an interquartile range (IQR) filter to remove noisy and outlier LiDAR points during 3D reconstruction. On KITTI dataset, TriBand-BEV attains 58.7/52.6/47.2 pedestrian BEV AP(%) for easy, moderate, and hard at 49 FPS on a single consumer GPU, surpassing Complex-YOLO, with gains of +12.6%, +7.5%, and +3.1%. Qualitative scenes show stable detection under occlusion. The pipeline is compact and ready for real time robotic deployment. Our source code is publicly available on GitHub.
Mohammad Khoshkdahan, Alexey Vinel
Karlsruhe Institute of Technology · Karlsruhe, Germany
Real-time 3D object detection is a critical component for the safe operation of autonomous driving systems and robotics. While LiDAR point clouds provide accurate spatial information, processing them efficiently remains a significant challenge. Traditional methods rely on complex 3D convolutions or anchor-based paradigms that struggle to balance detection accuracy with inference speed. In this paper, we propose PillarDETR, a novel end-to-end 3D object detection architecture that combines the efficiency of pillar-based LiDAR encoding with the representational power of modern 2D vision models. Specifically, PillarDETR replaces standard convolutional backbones with a Cross Stage Partial (CSP) network derived from YOLOv8, enabling richer feature extraction from pseudoimages. Furthermore, we discard conventional anchor-based or center-based detection heads in favor of a Real-Time Detection Transformer (RT-DETR) decoder. This hybrid design allows the network to capture global context and directly predict 3D bounding boxes without relying on non-maximum suppression (NMS). Extensive experiments on the KITTI and nuScenes benchmarks demonstrate that PillarDETR achieves a compelling trade-off between mean Average Precision (mAP) and inference latency. Our ablation studies confirm that integrating the YOLOv8 backbone and RT-DETR head yields substantial improvements over the PointPillars baseline, establishing PillarDETR as a highly effective solution for real-time 3D perception.
Smit Kadvani, Shriya Gumber, Kriti Faujdar +1
Independent Researcher · University of Southern California · Illinois Institute of Technology