U-Net-based mountain region offline path planning method

Abstract

The invention is applicable to the technical field of geographic information systems and path planning, and provides a U-Net-based mountain region offline path planning method, which is characterized in that images and laser point cloud data of a target region are acquired through an unmanned aerial vehicle, and a digital orthoimage (DOM) and a digital elevation model (DEM) are quickly made. Road extraction and vectorization are carried out on the DOM by adopting a U-Net deep learning algorithm, the road gradient is calculated by synchronously utilizing the DEM, and road center line vector data with width and gradient attributes are generated. And aiming at two working scenes of driving and walking, respectively adopting a cost weight improved RRT algorithm and a Dijkstra algorithm, and carrying out offline path planning based on the attribute road network. The method gets rid of dependence on a continuous network and a stable satellite positioning signal, solves the problem of path planning failure and inaccuracy caused by signal loss and old map data in a mountain environment, and realizes reliable and safe path planning in a non-signal area.

Description

Technical Field

[0001] This invention belongs to the field of geographic information systems and route planning technology, and particularly relates to an offline route planning method for mountainous areas based on U²-Net. Background Technology

[0002] Fieldwork is a core component of surveying, geological investigation, resource monitoring, scientific research, and emergency rescue. However, such operations are often located in remote mountainous areas, canyons, and dense forests, facing severe challenges such as sparse operator base station signals and satellite positioning signals (GPS / BeiDou) drifting or even failing due to terrain obstruction or vegetation attenuation. In this environment, conventional electronic navigation methods that rely on continuous network connectivity and stable satellite positioning cannot function properly, severely impacting operational efficiency and personnel safety.

[0003] To address the signal loss issue, the primary current solution is offline electronic navigation. This method requires operators to download a high-precision offline map of the work area (typically containing satellite imagery and pre-built road network vector data) to their mobile devices beforehand. In the absence of a network, the system relies on the terminal's own satellite positioning module to obtain its location and performs route planning and navigation based on the locally stored offline map data.

[0004] While offline electronic navigation alleviates reliance on real-time networks to some extent, it still has significant drawbacks: First, its navigation function depends entirely on the terminal's ability to continuously receive a sufficient number (usually at least four) of satellite signals for accurate positioning. In scenarios where signals are severely obstructed, such as valleys, dense forests, or underground tunnels, the positioning function will fail, leading to navigation interruption. Second, offline map data suffers from update lag, failing to reflect dynamic changes in the wilderness environment such as temporary roads, engineering facilities, landslides, and river course changes. Routes planned based on outdated maps may be unsuitable or even pose safety risks. Summary of the Invention

[0005] The purpose of this invention is to provide an offline path planning method for mountainous areas based on U²-Net, aiming to solve the problems existing in the background technology.

[0006] This invention is implemented as follows: an offline path planning method for mountainous areas based on U²-Net, which eliminates the dependence on continuous network connection and high-precision real-time satellite positioning, and can achieve safe and reliable path planning in offline mode, specifically including:

[0007] By equipping drones with optical lenses and laser point cloud (LIDAR) lenses, high-resolution images and point cloud data of the target area are collected along a preset flight path, providing a real-time and reliable data source for subsequent processing.

[0008] The collected data is preprocessed, aerial triangulated, and jointly adjusted to quickly generate digital orthophotos (DOM) and digital elevation models (DEM) for the region.

[0009] The U²-Net deep learning algorithm is used to extract and vectorize roads from the DOM, generating road centerline vector data that includes road width and slope attributes. The U²-Net algorithm adopts a nested dual-U-Net architecture and effectively addresses the challenges of extracting roads from long, narrow mountainous terrains with complex backgrounds through spatial attention (SAU) and feature attention (FAU) modules.

[0010] Based on the aforementioned road network with attribute information, offline path planning is performed using the RRT algorithm with cost weight optimization and the Dijkstra algorithm for both vehicular and pedestrian scenarios. The planning algorithms comprehensively consider road width, gradient, and length, aiming to optimize the overall travel cost and output a safe path that conforms to the characteristics of mountainous terrain.

[0011] The present invention provides an offline path planning method for mountainous areas based on U²-Net, which has the following beneficial effects:

[0012] This invention utilizes real-time UAV data acquisition and rapid modeling technology to dynamically generate digital orthophotos and elevation models reflecting the current terrain, overcoming the drawbacks of slow updates in traditional offline maps. Employing the U²-Net deep learning algorithm, it can extract road networks from complex mountain images with high precision and automatically, assigning key traffic attributes such as width and slope. Finally, based on this attributed road network, a cost-weighted path planning algorithm adapted to both vehicular and pedestrian scenarios is designed. Even in extreme environments with no base stations or satellite positioning signals, it can still plan optimized paths that balance safety and traffic efficiency. This method significantly improves the autonomous navigation capabilities, the rationality of path planning, and the overall safety level of mountain field operations. Attached Figure Description

[0013] Figure 1 The flowchart illustrates an offline path planning method for mountainous areas based on U²-Net, as provided by this invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0015] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0016] like Figure 1As shown, an offline path planning method for mountainous areas based on U²-Net is described, and its specific implementation steps are as follows:

[0017] Step S1: Acquisition of UAV imagery and LIDAR data.

[0018] A survey area analysis was conducted to clarify the UAV flight coverage boundary. Flight path parameters were designed based on the mountainous terrain characteristics, including: flight altitude controlled at 50-200m, forward overlap of 60%-80%, and lateral overlap of 30%-50%. This parameter design aims to ensure the integrity of data acquisition and the accuracy of subsequent modeling. After acquiring the preset flight path, it was transferred to the UAV flight control system. The UAV flew autonomously along the preset flight path, simultaneously activating its optical lens and LIDAR sensor to complete image and point cloud data acquisition. After the flight operation was completed, image data, POS attitude data, raw laser point cloud data, and inertial navigation data were exported to form the basic data source for subsequent processing.

[0019] Step S2: Quick creation of DOM and DEM.

[0020] First, data preprocessing is performed: optical image distortion correction technology is used to eliminate the effects of lens distortion; statistical filtering combined with morphological filtering algorithms is used to denoise the LiDAR point cloud, improving the quality of the raw data. Next, phase control points are acquired: existing control points in the survey area are prioritized; if control points are insufficient, corresponding feature points are matched in historical DOM and DEM and their coordinates are extracted as accuracy control benchmarks. Subsequently, aerial triangulation of the image and LiDAR point cloud coordinate calculation are performed, and joint adjustment optimization is conducted using the phase control point data. Finally, based on the calculation results, high-precision digital orthophoto images (DOM) and digital elevation models (DEM) are generated, providing texture information of ground features and topographic elevation information, respectively.

[0021] Step S3: U²-Net algorithm for road extraction and vectorization.

[0022] U²-Net is a deep learning-based image segmentation model that automatically and accurately extracts road networks from the DOM and assigns them accessibility attributes. Its core architecture employs a nested dual-U-Net path design, optimized for fine edge detection and small object segmentation. Through multi-scale feature fusion and attention mechanisms, it effectively addresses the challenges of extracting mountain roads from high-resolution remote sensing images.

[0023] The specific implementation process includes:

[0024] Preprocessing: The contrast of the DOM generated by S2 is enhanced to strengthen the grayscale continuity of the road and improve the distinction between the road and the background.

[0025] Data Preparation: Constructing the training dataset. A publicly available road dataset was selected, and the sample was expanded by combining local mountain road interpretation data. After cropping the images to a fixed size of 512×512 pixels, they were divided into training, validation, and test sets.

[0026] Feature extraction and segmentation: A multi-scale input strategy is used for model inference. The first-level U-Net processes the original DOM image, and the second-level U-Net processes the 1 / 2 resolution downsampled result of the original image to improve the model's global perception capability. Inside the algorithm, the Spatial Attention Unit (SAU) suppresses background interference such as vegetation and water by calculating spatial attention weights. The weight calculation formula is: M=σ(W2(W1[max(F),avg(F)])).

[0027] M: Spatial attention weight, used to weight the original feature F spatially, highlighting important regions and suppressing irrelevant regions.

[0028] F: Input feature map, usually represented as a three-dimensional tensor of H×W×C (height×width×number of channels).

[0029] [max(F), avg(F)]: Concatenates the results of max pooling and average pooling along the channel dimension. max(F) is the maximum pooling operation along the channel dimension, and avg(F) is the average pooling operation along the channel dimension;

[0030] W1: The weight matrix of the first fully connected layer. Dimensionality reduction, compression of image channels, and extraction of fusion features.

[0031] W2: The weight matrix of the second fully connected layer. It maps the number of channels back to the original dimension and outputs the inactive spatial attention weight map.

[0032] σ: Sigmoid activation function, which normalizes the output to 0~1 to obtain the final spatial attention weight map.

[0033] The Feature Attention Module (FAU) fuses local details and global semantic information through a channel attention mechanism to repair broken roads. The formula is: F... C (FAU(F)).

[0034] F C The original feature F was recalibrated by using attention weights, which highlighted important features and suppressed redundant information.

[0035] The input feature map is usually a three-dimensional tensor output by a convolutional layer, with dimensions C×H×W (number of channels × height × width).

[0036] : It is a normalization factor used to normalize the attention weights and prevent excessively large values ​​from causing training instability.

[0037] FAU(F): Weight coefficients calculated by the FeatureAttentionUnit. The importance weights for each feature location are calculated using a self-attention mechanism, which are then used to weight the original features, outputting a feature attention weight map with the same dimension as F.

[0038] Post-processing optimization: The road probability map output by the model is optimized. Morphological opening operation (erosion followed by dilation) is used to remove noise specks; the connectivity of the road network is optimized by combining the minimum spanning tree algorithm, and finally clean road raster data is output.

[0039] Vectorization and Attribute Assignment: The optimized road raster data is converted into vector polygons. After topology cleaning (removing small fragments and smoothing edges), the road centerline is extracted. The average polygon width is calculated and assigned to the corresponding road centerline. Simultaneously, elevation sampling (at 10-meter intervals) is performed on the DEM generated by S2 using the road centerline. The road slope is obtained by calculating the elevation difference between adjacent sampling points, and after equalization, it is assigned to the centerline. Finally, road centerline vector data containing width (representing accessibility) and slope (representing risk level) attributes is formed.

[0040] Step S4: Offline route planning.

[0041] This step is based on the road centerline vector network with attributes generated by S3, and performs optimal path search for different operation modes.

[0042] Vehicle Path Planning: Traditional RRT algorithms are suitable for obstacle avoidance path searching in unstructured environments such as open areas and obstacle zones, but are difficult to adapt to structured road networks in mountainous terrain. This invention improves the cost weighting of the RRT algorithm, defining the "optimal path" as the one with the best overall driving cost (considering traversability, fuel consumption, and time cost). The core improvement is the construction of a three-dimensional cost weighting system based on width, gradient, and length. Regarding road width W... width The narrower the width, the lower the traffic efficiency and the greater the cost weight. For a standard two-lane road (≥7m), the weight coefficient is 1 (optimal). For single-lane roads (<3.5m), the weight coefficient increases accordingly. Roads <3m are directly marked as impassable. Regarding road slope (Wslope), the greater the absolute value of the slope, the higher the driving risk and fuel consumption, and the greater the weight coefficient. For flat slopes (≤3%), the weight coefficient is 1. For uphill scenarios, the penalty coefficient k=0.3; for downhill scenarios, k=0.2. Slopes >15% are marked as impassable. slope=1+k*|slope value|. Road length is the basic cost factor, taken as the actual length L of the vector road segment. The comprehensive cost calculation formula is: Cost=L×W width ×W slope (W) width W is the width weighting coefficient. slope (This refers to the slope weighting coefficient).

[0043] The algorithm execution flow is as follows:

[0044] ①Structural modeling of the road network. The vector road network is abstracted into a directed graph G=(V,E), where V is the set of nodes such as road intersections and the start and end points of operations, and E is the set of directed edges of road segments (clearly defining one-way / two-way traffic rules). Each edge is assigned a comprehensive cost and impassable edges are removed.

[0045] ② Initialize the tree structure. Construct a tree T={S} with the starting point S as the root node, define the expansion radius r (take the average adjacency distance of the nodes), and limit the search to only the adjacent nodes.

[0046] ③ Constrained random sampling. Candidate point q rand Forced to fall within the node set V, introduce a 10% probability of target bias sampling (direct sampling of the endpoint G) to improve efficiency.

[0047] ④ Find the tree T containing the element q rand The node q with the minimum overall cost near .

[0048] ⑤ Road network constraint extension. If q rand With q near If a walkable connected edge exists, then generate a new node q. new= q rand And add it to tree T.

[0049] ⑥ Parent node reselection optimization. In q new Filtering neighbor nodes q within the extended radius near_neighbors Calculate the distance from each node to q via the starting point. new The total cost is used to select the optimal parent node, update the tree structure, and prune the branches.

[0050] ⑦ Backtracking to the destination. Backtracking tree T from the destination G back to the starting point S yields the vehicle path with the minimum overall cost, which is the optimal planned path.

[0051] Pedestrian path planning: Utilizing the classic Dijkstra algorithm, adapted to pedestrian network planning requirements with width and slope constraints, the core is the construction of a cost-weight model based on pedestrian operation characteristics. The cost factor design logic is as follows: when the road width is <2m, pedestrian avoidance space is insufficient and safety is poor; the weight coefficient increases as the width decreases. For a comfortable walking width ≥4m, the weight coefficient is set to 1 (base value). The calculation formula is W. width=max(1,4 / actual width), width < 1.5m is marked as difficult to walk on. Regarding road slope, uphill slopes increase physical exertion, while downhill slopes pose a risk of slipping. The larger the absolute value of the slope, the higher the weight. For flat slopes (≤3%), the coefficient is 1. The calculation formula is W. slope =1 + k * |slope value| (k=0.2 for uphill, k=0.15 for downhill), slopes > 30% are marked as impassable. Road length is taken as the actual length L, and the comprehensive cost formula is the same as for traffic planning: Cost = L × W width ×W slope (W) width W is the width weighting coefficient. slope (This refers to the slope weighting coefficient).

[0052] The algorithm execution flow is as follows:

[0053] ① Road network preprocessing. Traverse the vector road network and remove impassable road sections with a slope greater than 30% and a width less than 1.5m.

[0054] ② Initialize the cost matrix. Assume the initial cost from the starting point to all nodes is infinite, and the cost of the starting point itself is 0. Construct a set of unvisited nodes.

[0055] ③ Iterative search. Each time, select the node with the smallest cost from the unvisited nodes as the current node, traverse its adjacent nodes and calculate the total cost to reach the adjacent nodes through the current node.

[0056] ④ Cost update. If the newly calculated total cost is less than the current recorded cost of the adjacent nodes, then update the cost and record the predecessor node.

[0057] ⑤ Repeat the above steps until the endpoint or all nodes are traversed. Backtrack from the endpoint to the starting point to obtain the optimal walking path plan.

[0058] Through the above steps, this invention can complete offline route planning that conforms to actual travel conditions in mountainous environments without network or reliable satellite positioning, based on real-time collection and processing of local geographic data.

[0059] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for offline path planning in mountainous areas based on U²-Net, characterized in that, The offline path planning method for mountainous areas based on U²-Net includes: S1. UAV imagery and LIDAR data acquisition: Analyze the survey area, clarify the UAV flight coverage boundary, design flight path parameters in combination with mountainous terrain features, obtain the preset flight path and transmit it to the UAV, and simultaneously collect image data and point cloud data during the UAV's flight along the preset flight path. S2, DOM and DEM Rapid Production: Preprocess the acquired image data and point cloud data, obtain phase control points, calculate the coordinates of the aerial triangulation of the image and the LiDAR point cloud, perform joint adjustment optimization in combination with the phase control point data, and generate digital orthophoto image DOM and digital elevation model DEM based on the calculation results. S3 and U²-Net algorithm for road extraction and vectorization: The U²-Net algorithm is used to extract and vectorize roads from digital orthophoto DOM to obtain road centerline vector data with width and slope attributes; S4. Offline route planning: Route planning is performed based on road centerline vector data containing width and slope attributes.

2. The offline path planning method for mountainous areas based on U²-Net according to claim 1, characterized in that, In S1, the route parameters include: flight altitude 50-200m, heading overlap 60%-80%, and lateral overlap 30%-50%.

3. The offline path planning method for mountainous areas based on U²-Net according to claim 1, characterized in that, In S2, the acquisition of phase control points specifically involves: prioritizing the use of existing control points in the survey area; if there are insufficient control points, matching corresponding feature points and extracting their coordinates from historical digital orthophoto images (DOM) and digital elevation models (DEM).

4. The offline path planning method for mountainous areas based on U²-Net according to claim 1, characterized in that, S3 specifically includes: The digital orthophoto DOM is preprocessed to enhance contrast, and a training dataset is constructed. A multi-scale input strategy is used for feature extraction and segmentation. The first-level U-Net processes the original image, and the second-level U-Net processes the 1 / 2 resolution downsampled image. Post-processing optimization is performed, morphological opening operation is used to remove noise specks, and the minimum spanning tree algorithm is combined to optimize road connectivity, outputting clean road raster data. Vectorization and attribute assignment are performed to convert the raster roads into vector polygons. After topology cleaning, the road centerlines are extracted, the average polygon width is calculated and assigned to the centerlines, and the centerlines are used to perform elevation sampling on the digital elevation model (DEM) to calculate the slope and assign it to the centerlines.

5. The offline path planning method for mountainous areas based on U²-Net according to claim 4, characterized in that, In the feature extraction and segmentation steps of S3, the SAU module in the U²-Net algorithm suppresses interference by calculating spatial attention weights. The formula for calculating the spatial attention weights is as follows: M=σ(W2(W1[max(F),avg(F)])); in: M: Spatial attention weight, used to weight the original features F in terms of spatial dimensions; F: Input feature map, represented as a three-dimensional tensor of height × width × number of channels; [max(F), avg(F)]: This indicates that the results of max pooling and average pooling are concatenated along the channel dimension. max(F) is the maximum pooling operation along the channel dimension, and avg(F) is the average pooling operation along the channel dimension. W1: Weight matrix of the first fully connected layer; W2: Weight matrix of the second fully connected layer; σ: Sigmoid activation function, which normalizes the output to 0~1 to obtain the final spatial attention weight map.

6. The offline path planning method for mountainous areas based on U²-Net according to claim 4, characterized in that, In S3, the specific steps for calculating the slope using centerline-based elevation sampling of the digital elevation model (DEM) are as follows: The slope is obtained by sampling the DEM at 10-meter intervals and calculating the difference between adjacent elevations.

7. The offline path planning method for mountainous areas based on U²-Net according to claim 1, characterized in that, In S4, the offline path planning includes vehicle path planning and pedestrian path planning; The vehicle route planning uses the cost-weighted improved RRT algorithm, and the pedestrian route planning uses the Dijkstra algorithm; The combined cost calculation formula for both vehicle route planning and pedestrian route planning is: Cost = L × W width ×W slope Where L is the actual length of the road segment, and W width W is the width weighting coefficient. slope W is the slope weighting coefficient. slope =1+k*|slope value|, where k is the penalty coefficient for uphill scenarios.

Images