In this paper, we propose a road surface segmentation technique which is accurate and suitable for hardware implementation. The roadsurface is segmented by detecting the boundary between road and obstacle, based on t...
详细信息
In this paper, we propose a road surface segmentation technique which is accurate and suitable for hardware implementation. The roadsurface is segmented by detecting the boundary between road and obstacle, based on the disparity histogram which we define as VLDH (Vertically Local Disparity Histogram). On each pixel of disparity image, VLDH is computed from the disparities of vertically local neighbourhood pixels. The major advantage is the feasibility of the pipeline processing on image processing hardware for stereo camera. The direct advantage on processing time is confirmed based on implementation into FPGA. Experimental result also shows that the accuracy of the proposed method is better than the conventional method.
This paper proposes a novel algorithm for extracting street light poles from vehicleborne mobile light detection and ranging (LiDAR) point-clouds. First, the algorithm rapidly detects curb-lines and segments a point-c...
详细信息
This paper proposes a novel algorithm for extracting street light poles from vehicleborne mobile light detection and ranging (LiDAR) point-clouds. First, the algorithm rapidly detects curb-lines and segments a point-cloud into road and nonroadsurface points based on trajectory data recorded by the integrated position and orientation system onboard the vehicle. Second, the algorithm accurately extracts street light poles from the segmented nonroadsurface points using a novel pairwise 3-D shape context. The proposed algorithm is tested on a set of point-clouds acquired by a RIEGL VMX-450 mobile LiDAR system. The results show that roadsurfaces are correctly segmented, and street light poles are robustly extracted with a completeness exceeding 99%, a correctness exceeding 97%, and a quality exceeding 96%, thereby demonstrating the efficiency and feasibility of the proposed algorithm to segment roadsurfaces and extract street light poles from huge volumes of mobile LiDAR point-clouds.
We propose a novel adaptation method for generalizing roadsegmentation to novel weather, lighting or viewing geometries. The method assumes a source domain consisting of an ensemble of labeled training datasets and a...
详细信息
ISBN:
(纸本)9781728198910
We propose a novel adaptation method for generalizing roadsegmentation to novel weather, lighting or viewing geometries. The method assumes a source domain consisting of an ensemble of labeled training datasets and an unlabeled target test dataset that deviates substantially from the training ensemble. The training dataset is used to compile a geometry-anchored prior over the road pixel locations and to train a fully-convolutional network roadsegmentation system. At inference, a probabilistic Houghing method is used to detect line intersections in the test image and thereby estimate the vanishing point of the road, thus anchoring the learned geometric prior. This prior is then used to extract high confidence road and background regions which serve as surrogate ground truth to adapt the network to the target domain. Leave-one-out evaluation across five diverse roadsegmentation datasets demonstrates substantial improvement in generalization across changes in viewing geometry and weather conditions, yielding results that are on average comparable and in some cases superior to a more complex GAN-based domain adaptation approach. These results demonstrate the potential for classical computer vision methods to guide adaptation of supervised machine learning algorithms, leading to improved generalization across domains.
Point-cloud semantic segmentation is a visual task essential for agricultural robots to comprehend natural agroforestry environments. However, owing to the extremely large amount of point-cloud data in agroforestry en...
详细信息
Point-cloud semantic segmentation is a visual task essential for agricultural robots to comprehend natural agroforestry environments. However, owing to the extremely large amount of point-cloud data in agroforestry environments, learning effective features for semantic segmentation from large-scale point clouds is challenging. Therefore, to address this issue and achieve accurate semantic segmentation of different types of road-surface point clouds in large-scale agroforestry environments, this study proposes a point-cloud semantic segmentation network framework based on double-distance self-attention. First, a point-cloud local feature enhancement module is proposed. This module primarily extends the receptive field and enhances the generalizability of multidimensional features by incorporating reflection intensity information and a spatial feature- encoding block that is enhanced with contextual semantic information. Second, we introduce a dual-distance attention pooling (DDAPS) block based on the self-attention mechanism. This block initially learns the feature representation of the local neighborhood of each point through the self-attention mechanism. Then, it uses the DDAPS block to aggregate more discriminative local neighborhood point features. Finally, extensive experimental results on large-scale point-cloud datasets, SemanticKITTI and RELLIS-3D, demonstrate that our algorithm outperforms similar algorithms in large-scale agroforestry environments.
Nowadays, precise and up-to-date maps of road are of great significance in an extensive series of applications. However, it automatically extracts the roadsurfaces from high-resolution remote sensed images which will...
详细信息
Nowadays, precise and up-to-date maps of road are of great significance in an extensive series of applications. However, it automatically extracts the roadsurfaces from high-resolution remote sensed images which will remain as a demanding issue owing to the occlusion of buildings, trees, and intricate backgrounds. In order to address these issues, a robust Gradient Descent Sea Lion Optimization-based U-Net (GDSLO-based U-Net) is developed in this research work for road outward extraction from High Resolution (HR) sensing images. The developed GDSLO algorithm is newly devised by the incorporation of Stochastic Gradient Descent (SGD) and Sea Lion Optimization Algorithm (SLnO) algorithm. Input image is pre-processed and U-Net is employed in roadsegmentation phase for extracting the roadsurfaces. Meanwhile, training data of U-Net has to be done by using the GDSLO optimization algorithm. Once roadsegmentation is done, road edge detection and road centerline detection is performed using Fully Convolutional Network (FCN). However, the developed GDSLO-based U-Net method achieved superior performance by containing the estimation criteria, including precision, recall, and F1-measure through highest rate of 0.887, 0.930, and 0.809, respectively.
暂无评论