In recent years, the application of China UAV drone technology in civil infrastructure inspection has witnessed remarkable growth. Our research team has been dedicated to developing a safe and adaptive trajectory planning framework for China UAV drone operations in complex bridge environments. The challenges posed by dense obstacle distributions, high-dimensional non-convex solution spaces, and stringent safety constraints necessitate innovative approaches that integrate lightweight environment modeling with advanced meta-heuristic optimization algorithms. This paper presents our comprehensive study on achieving autonomous and collision-free navigation for China UAV drone during intricate bridge inspection tasks.
1. Introduction
The inspection of large-span bridges using China UAV drone platforms has emerged as a transformative alternative to traditional manual inspection methods. However, the autonomous path planning for China UAV drone in such environments remains a formidable challenge due to the presence of densely arranged cables, tower pylons, and beam structures that create narrow passageways and shadowed regions. The core requirements for China UAV drone trajectory planning include obstacle avoidance, energy efficiency, kinematic feasibility, and real-time computational performance.
Existing studies on China UAV drone path planning have explored various meta-heuristic algorithms, including particle swarm optimization, grey wolf optimizer, and genetic algorithms. Nevertheless, these methods often suffer from premature convergence and inadequate exploration capabilities when dealing with the high-dimensional, multi-constrained optimization problems typical of bridge inspection scenarios. To address these limitations, we propose a novel enhanced algorithm that integrates Latin hypercube sampling with a sparrow search mechanism into the artificial lemming algorithm framework, specifically designed for China UAV drone trajectory generation.
The key contributions of our work can be summarized as follows:
| Contribution | Description |
|---|---|
| Lightweight environment modeling | A semantic filtering approach based on IFC standards for extracting critical bridge components and constructing safe observation spaces via normal vector extrapolation |
| Enhanced optimization algorithm | An LSS-ALA framework combining Latin hypercube sampling, sparrow search vigilance mechanism, and nonlinear adaptive weight strategy for superior global exploration and local exploitation |
| Multi-constraint trajectory generation | A comprehensive objective function incorporating flight distance, energy consumption, turning angle limitations, and obstacle avoidance, smoothed via cubic B-spline interpolation |
| Comprehensive validation | Extensive testing on CEC2022 benchmark functions and a realistic long-span cable-stayed bridge model demonstrating significant improvements in convergence accuracy, robustness, and trajectory smoothness |
2. Lightweight Environment Modeling Based on BIM Semantics
The first critical step in enabling safe China UAV drone navigation is the construction of an accurate yet computationally efficient environmental representation. Raw building information models contain dense geometric data, including reinforcement bars and piping networks, which can generate millions of triangular facets. Directly processing such data imposes prohibitive computational burdens on real-time path planning algorithms. We developed a semantic filtering methodology based on the IFC 4×3 standard data dictionary to address this issue.
2.1 Semantic Filtering and Geometric Simplification
Our semantic filtering pipeline follows a negative removal and positive retention strategy. We extract structural load-bearing components such as cable towers (IfcColumn), main girders (IfcBeam), and stay cables (IfcCableSegment) while eliminating non-structural elements including guardrails (IfcRailing), lighting fixtures (IfcLightFixture), and drainage pipes (IfcPipeSegment) through Boolean difference operations. The geometric resolution offset is set to 5 mm to preserve surface curvature fidelity for tower pylons. Each component category is exported as a separate STL file, achieving effective decoupling of spatial geometry from underlying attribute semantics.
The lightweight processing pipeline is illustrated through the following data flow:
The complete processing chain from Revit modeling environment to Matlab simulation platform comprises:
- Stage 1 (Revit): Semantic view filtering removes redundant components; layered export of independent STL files
- Stage 2 (Matlab): STL mesh parsing extracts surface vertex sets and normal vectors
To quantify the computational load reduction achieved by our BIM semantic lightweight technique, we conducted comparative analysis between the original complete BIM model and the lightweight model. The results demonstrate substantial improvements:
| Metric | Original BIM Model | Lightweight Model | Reduction (%) |
|---|---|---|---|
| Geometric facet count | 2,850,000 | 99,750 | 96.5% |
| Single collision detection time (ms) | 145.2 | 4.8 | 96.7% |
| Algorithm total runtime (s) | 125.6 | 5.1 | 95.9% |
| Memory footprint (MB) | 1,240 | 86 | 93.1% |
2.2 Safe Observation Space Construction via Normal Extrapolation
We constructed a safe observation point set by extrapolating the surface normal vectors of bridge components outward, combined with adaptive discrete sampling based on component type. This approach transforms the complex obstacle avoidance problem into a static boundary constraint optimization.
The three-dimensional spatial generation formula for characteristic observation points is:
$$P_{obs} = P_{surf} + n \cdot d_{safe}$$
where Pobs represents the spatial coordinates of the safe observation point for China UAV drone, Psurf denotes the coordinates of feature points on the bridge component surface, n is the outward unit normal vector at the feature point, and dsafe is the preset safe inspection distance derived from the physical dimensions of China UAV drone and flight error margins:
$$d_{safe} = R_{UAV} + \delta_{GPS} + \sigma_{wind} + d_{buffer}$$
where RUAV = 0.45 m (rotor expansion radius of China UAV drone), δGPS = 0.1 m (RTK positioning drift error), σwind = 0.45 m (maximum trajectory offset under gust disturbance), and dbuffer = 1.0 m (camera obstacle avoidance blind zone). Based on these values, we set dsafe = 2.0 m.
The inspection space Ω is partitioned into three regions: obstacle region Ωobs, safety buffer zone Ωsafe, and feasible flight region Ωfree. After importing the lightweight model into the simulation platform, we implemented adaptive sampling strategies for different component types:
- Dense sampling for complex topological regions with step size Smin = 1 m to ensure complete coverage of critical areas
- Sparse sampling for flat bridge deck areas with step size Smax = 10 m to reduce point set cardinality
This adaptive strategy significantly reduces the number of observation points and algorithm search dimensions while maintaining inspection completeness for key regions, providing boundary constraints for the subsequent LSS-ALA path optimization for China UAV drone.
3. Mathematical Characterization of Multi-dimensional Inspection Constraints
In practical China UAV drone operations, the trajectory planning must satisfy physical performance limitations. We constructed kinematic and energy consumption constraint models to ensure both flight safety and mission feasibility.
3.1 Kinematic Constraints
To maintain flight stability and prevent gimbal camera from losing target due to剧烈的 attitude changes, we restrict the turning angle between adjacent flight segments. Let the heading vectors be v1 = Pi − Pi−1 and v2 = Pi+1 − Pi. The constraint is expressed as:
$$\theta_i = \arccos\left(\frac{\mathbf{v}_1 \cdot \mathbf{v}_2}{|\mathbf{v}_1| \cdot |\mathbf{v}_2|}\right) \leq \theta_{max}$$
where θi is the spatial turning angle of China UAV drone at the i-th waypoint, and θmax is the maximum allowable turning angle. Exceeding this threshold would cause motion blur in captured images and increase the risk of collision during sharp maneuvers.
3.2 Energy Consumption Constraints
Considering the significantly higher energy consumption of rotary-wing China UAV drone during vertical climbing compared to horizontal cruising, we established a weighted energy consumption model. The total estimated energy consumption Etotal for a single inspection mission must not exceed the safety threshold Emax:
$$E_{total} = \sum_{i=1}^{N-1} \left(k_h \cdot D_{xy}(i,i+1) + k_v \cdot |z_{i+1} – z_i|\right) \leq E_{max}$$
where kh and kv are the energy consumption coefficients for horizontal flight and vertical climbing, respectively; Dxy(i,i+1) is the horizontal flight distance between adjacent waypoints; zi+1 and zi are the altitude values; and N is the total number of waypoints in the planned path.
3.3 Comprehensive Objective Function
We constructed a composite objective function with dynamic penalty terms to obtain the shortest inspection path satisfying kinematic and energy constraints:
$$\min F(\mathbf{X}) = \sum_{i=1}^{N} D(\mathbf{P}_i, \mathbf{P}_{i+1}) + \omega_1 \sum_{i=1}^{N} \Psi(\theta_i) + \omega_2 \Phi(E_{total})$$
where F(X) is the comprehensive objective function; the first term represents the basic optimization goal of minimizing total flight distance; Ψ(θi) and Φ(Etotal) are penalty functions for turning angle violation and energy overrun, respectively; and ω1 and ω2 are corresponding penalty weight coefficients. The dynamic penalty weight strategy effectively suppresses China UAV drone trajectories from entering infeasible regions during early optimization stages.
4. LSS-ALA: Enhanced Artificial Lemming Algorithm for China UAV Drone
4.1 Standard Artificial Lemming Algorithm
The standard artificial lemming algorithm simulates the collective behavior of lemming populations through mathematical mapping. During optimization, the transition between global exploration and local exploitation is regulated by an energy factor Fe. When |Fe| > 1, individuals perform long-range migration to expand the search range; when |Fe| ≤ 1, the population transitions to burrowing or foraging states for intensified local exploitation. The energy factor undergoes nonlinear decay with iteration count t:
$$F_e(t) = 4 \cdot \arctan\left(1 – \frac{t}{T_{max}}\right) \times \ln\left(\frac{1}{R}\right)$$
where R is a random number uniformly distributed in (0, 1), and Tmax is the maximum number of iterations.
4.2 LSS-ALA Improvement Strategies
The conventional ALA suffers from uneven initial population distribution and susceptibility to local optima in complex obstacle environments. We developed the LSS-ALA collaborative model with three key improvements.
4.2.1 Latin Hypercube Sampling Initialization
Given the narrow feasible solution space in bridge environments, we replaced pseudo-random initialization with Latin hypercube sampling. By uniformly stratifying the value range of each decision variable dimension, LHS ensures traversability and uniformity of the initial population in the solution space, reducing the algorithm’s sensitivity to initial values. The sampling formula for the j-th dimension component of the i-th individual is:
$$X_{ij} = B_{l,j} + \frac{B_{u,j} – B_{l,j}}{N} \times (P_{ji} – R)$$
where Bu,j and Bl,j are the upper and lower bounds of the j-th dimension; Pji is the i-th element in the random sequence of 1 to N corresponding to the j-th dimension. For the three-dimensional inspection space, the number of LHS sampling intervals matches the initial population size (N = 30), with each dimension’s feasible domain divided into 30 equal-probability intervals, each extracted exactly once. This approach avoids excessive subdivision that would increase time cost while ensuring coverage of narrow regions such as tower back sides and cable intersection zones.
The comparative analysis between pseudo-random and LHS initialization reveals distinct advantages. The conventional pseudo-random scheme tends to exhibit local clustering with obvious exploration blind spots, while LHS enables the initial population to evenly cover the solution domain with minimal overlap and gaps in projections. This uniform distribution effectively compensates for early spatial exploration deficiencies, providing a solid foundation of population diversity for global optimization of China UAV drone trajectories.
4.2.2 Embedded Sparrow Search Alert Mechanism
To address the tendency of conventional ALA to陷入 local optima in later iterations due to energy factor decay, we introduced the vigilance mechanism from the sparrow search algorithm to reconstruct the position update strategy. By monitoring population status, when local stagnation warnings are triggered, suboptimal individuals perform directed transitions guided by the global optimal position supplemented with Lévy flight operators, enhancing the algorithm’s ability to escape local optima:
$$\mathbf{X}_i^{t+1} = \mathbf{X}_{best}^t + A(t) \cdot |\mathbf{X}_i^t – \mathbf{X}_{best}^t| \cdot L \cdot \text{sign}(R – 0.5)$$
where Xbestt is the global extremum coordinate at iteration t, and L is the Lévy flight operator responsible for expanding transition randomness:
$$L = \frac{u}{|v|^{1/\beta}}$$
where β = 1.5 is a constant; u ~ N(0, σu2), v ~ N(0, σv2), with:
$$\sigma_u = \left[ \frac{\Gamma(1+\beta) \cdot \sin(\pi\beta/2)}{\Gamma((1+\beta)/2) \cdot \beta \cdot 2^{(\beta-1)/2}} \right]^{1/\beta}, \quad \sigma_v = 1$$
4.2.3 Hybrid Strategy Adaptive Alternation
To balance lemming behavior and sparrow alert mechanism, we set a probability threshold Pswitch. During the iteration process, individual behavior follows the criterion:
$$\mathbf{X}_i^{t+1} =
\begin{cases}
S_{SSA}, & R < P_{switch} \\
S_{ALA}, & \text{otherwise}
\end{cases}$$
where SSSA and SALA correspond to the sparrow alert mechanism constructed in Equation (8) and the traditional lemming iteration rule, respectively, with the switching controlled by threshold Pswitch. This adaptive alternation ensures that China UAV drone trajectory optimization benefits from both global exploration during early stages and refined local exploitation during later stages.
5. Trajectory Smoothing and Algorithm Complexity Analysis
5.1 Cubic B-Spline Trajectory Smoothing
The paths generated by optimization algorithms are discrete polylines that cannot be directly used for flight control of China UAV drone. We employ cubic B-spline interpolation as a post-processing step, leveraging its C2 continuity to eliminate geometric sharp corners and generate a continuous smooth trajectory satisfying the dynamic constraints of China UAV drone:
$$\mathbf{S}(t) = \sum_{i=0}^{n} \mathbf{P}_i \cdot N_{i,k}(t)$$
where S(t) represents the smoothed continuous trajectory coordinates; Pi are the discrete waypoints generated by the algorithm (control vertices); and Ni,k(t) are the k-th degree B-spline basis functions with k = 3. The C2 continuity ensures that the trajectory curvature is continuous, enabling China UAV drone to execute smooth turns without abrupt acceleration changes.
5.2 Computational Complexity Analysis
The introduction of hybrid strategies in LSS-ALA does not increase the core loop nesting depth of the algorithm. LHS is executed only once during the initialization phase with complexity O(ND), where N is the population size and D is the problem dimension. The SSA position update mechanism is triggered based on probability and only replaces part of the ALA update steps without introducing additional computational overhead. Therefore, the single-iteration complexity of LSS-ALA remains at O(ND), and the total time complexity remains O(TND) where T is the maximum number of iterations. This improvement satisfies real-time requirements for engineering applications without significantly increasing computational time.
The key parameters of LSS-ALA and penalty weight coefficients were determined through control variable experiments and pre-testing on representative CEC2022 functions. Sensitivity analysis results indicate that when the discoverer proportion is 0.2 and the vigilance threshold is 0.8, the algorithm achieves optimal dynamic balance between global exploration and local exploitation.
| Algorithm | Parameter | Value |
|---|---|---|
| LSS-ALA | Discoverer proportion | 0.2 |
| Vigilance threshold | 0.8 | |
| Inertia weight range | [2, 0] | |
| ALA | Behavior switch threshold | 0.5 |
| GWO | Convergence factor a | [2, 0] |
| SABO | Exploration step Rstep | 0.1 |
| CPO | Defense mechanism Pdefense | 0.1 |
| TLCO | Teaching factor | Randomly selected 1 or 2 |
6. Experimental Validation
6.1 Benchmark Function Testing on CEC2022 Suite
To validate the theoretical performance of LSS-ALA, we selected the CEC2022 benchmark suite covering four categories (unimodal, basic, hybrid, and composition) with a total of 12 test functions. The experimental platform employed an Intel Core i7-12700H CPU (2.30 GHz) with 16 GB of RAM, running Matlab R2024a. Comparison algorithms included crested porcupine optimization, termite life cycle optimization, grey wolf optimizer, subtraction-average-based optimizer, and the baseline artificial lemming algorithm. Experimental parameters were uniformly set as: population size N = 30, maximum iterations Tmax = 500, and dimension D = 20.
To eliminate random errors, each algorithm was independently executed 30 times on each test function. The comprehensive optimization results demonstrate that LSS-ALA achieves the optimal values on the vast majority of functions, particularly excelling in the hybrid function F6 by leveraging the hybrid strategy to escape local optima, exhibiting strong robustness.
| Function | Type | LSS-ALA Mean | LSS-ALA Std | Best Competitor Mean | Best Competitor Std |
|---|---|---|---|---|---|
| F1 | Unimodal | 1.52e-08 | 2.31e-09 | 3.87e-06 | 5.42e-07 |
| F2 | Unimodal | 4.86e-07 | 6.73e-08 | 2.15e-05 | 8.94e-06 |
| F3 | Basic | 8.33e-06 | 1.22e-06 | 4.71e-04 | 3.56e-05 |
| F4 | Basic | 1.97e-04 | 2.84e-05 | 6.32e-03 | 7.18e-04 |
| F5 | Basic | 2.61e-05 | 3.95e-06 | 8.14e-03 | 1.05e-03 |
| F6 | Hybrid | 5.73e-03 | 8.12e-04 | 9.85e-01 | 1.47e-01 |
| F7 | Hybrid | 1.24e-02 | 1.76e-03 | 2.36e+00 | 4.52e-01 |
| F8 | Hybrid | 3.45e-03 | 4.89e-04 | 7.24e-01 | 8.93e-02 |
| F9 | Hybrid | 8.91e-04 | 1.23e-04 | 5.16e-02 | 6.78e-03 |
| F10 | Composition | 2.67e-02 | 3.41e-03 | 1.45e+01 | 2.86e+00 |
| F11 | Composition | 1.83e-01 | 2.56e-02 | 8.73e+00 | 1.34e+00 |
| F12 | Composition | 4.56e-02 | 6.23e-03 | 3.28e+00 | 5.67e-01 |
The average single-run time of LSS-ALA on the CEC2022 test set is 0.09 s, compared to 0.07 s for the baseline ALA, demonstrating that the hybrid strategy does not introduce significant computational overhead while achieving superior optimization accuracy. Statistical analysis using the Wilcoxon rank-sum test confirmed that in most scenarios, the p-values obtained by LSS-ALA are less than 0.05, verifying the statistical significance of the algorithm.
6.2 Convergence Behavior Analysis
The convergence dynamics of LSS-ALA across the full optimization cycle exhibit distinct advantages. For unimodal function F1 and composition function F11, LSS-ALA demonstrates rapid vertical descent in the early iterations (first 50 generations), primarily attributed to the LHS initialization strategy. Compared to CPO and SABO with pseudo-random initialization, the initial population of LSS-ALA is more uniformly distributed in the solution space, reducing the time cost of early blind search.
When processing high-complexity hybrid and composition functions such as F6 and F12, comparison algorithms including GWO and TLCO frequently exhibit convergence stagnation or stepwise slow descent, easily陷入 local optima. In contrast, LSS-ALA maintains high convergence efficiency from early stages, with fitness values rapidly decreasing and stabilizing within fewer iterations. The integration of the sparrow search “leader-follower” mechanism effectively enhances the global detection capability of the population, enabling leader individuals to perform random transitions that help the population escape local attractor basins.
6.3 Bridge Inspection Path Planning Simulation
To validate the applicability of our algorithm in realistic complex scenarios, we constructed a three-dimensional test environment based on an actual in-service long-span three-tower double-cable-plane cable-stayed bridge model with a total length of 600 m, main span of 320 m, and tower height of 110 m. The bridge deck features densely arranged stay cables at inclination angles ranging from 30° to 60°. Unlike conventional global inspection, our experiment focused on targeted defect inspection of high-risk components including tower anchorage zones and main girder bottom plate joints. The target points are predominantly located in non-convex spaces where multiple component intersections create narrow passageways with minimum clearance below 4 m, further constrained by safety buffer zones, resulting in confined feasible flight regions characterized by high density, multiple blind spots, and strongly coupled obstacle features.
The comprehensive performance comparison of path planning results across six algorithms reveals significant differences:
| Algorithm | Lbest (m) | Lavg (m) | Std (m) | θtotal (°) |
|---|---|---|---|---|
| LSS-ALA | 1,996.15 | 2,060.32 | 21.36 | 1,245.6 |
| CPO | 2,105.42 | 2,198.50 | 85.12 | 1,350.8 |
| TLCO | 2,156.88 | 2,245.10 | 110.45 | 1,412.5 |
| ALA | 2,285.30 | 2,410.65 | 196.22 | 1,520.4 |
| GWO | 2,412.50 | 2,655.80 | 305.14 | 1,650.9 |
| SABO | 2,530.12 | 2,840.25 | 412.30 | 1,720.5 |
The results demonstrate that LSS-ALA achieves the shortest safe path length of 1,996.15 m, representing a 17.2% reduction compared to GWO. This substantial improvement indicates that our enhancement strategy effectively boosts the global optimization capability in complex high-dimensional spaces while reducing the risk of falling into local optima. In terms of trajectory smoothness and engineering safety, LSS-ALA achieves the best control effect with a cumulative turning angle of 1,245.6°, which is 27.6% lower than SABO. Limiting excessive heading changes is crucial for practical engineering applications, as smooth trajectories reduce motion blur caused by airframe directional changes, ensure image acquisition quality, and reduce additional energy consumption from frequent motor acceleration and deceleration.
The standard deviation across 30 independent runs is merely 21.36, representing the smallest dispersion among all comparison algorithms. This indicates that our algorithm exhibits excellent optimization stability and robustness when solving complex bridge inspection paths with dense obstacles.
7. Discussion and Future Work
Our proposed LSS-ALA framework demonstrates significant advantages for China UAV drone trajectory planning in complex bridge environments. The integration of BIM semantic lightweight modeling with enhanced meta-heuristic optimization creates a synergistic effect that addresses both computational efficiency and solution quality. The semantic filtering approach reduces geometric complexity by 96.5%, enabling real-time collision detection that is essential for practical China UAV drone operations. The LSS-ALA algorithm, with its Latin hypercube sampling initialization and sparrow search vigilance mechanism, effectively overcomes the premature convergence issue that plagues conventional meta-heuristic algorithms in high-dimensional constrained spaces.
However, the current planning system has not yet considered wind field interference in bridge regions, such as gusts and crosswinds. Under extreme wind loads, the safety margin of static boundary constraints may be insufficient, and the online re-planning capability of our algorithm under wind perturbation requires further validation. Future research directions include:
- Developing wind field modeling in complex bridge environments using computational fluid dynamics simulations to characterize spatial and temporal wind patterns
- Exploring real-time observation space updating mechanisms based on dynamic wind direction prediction to enhance operational reliability under adverse meteorological conditions
- Integrating reinforcement learning approaches for adaptive parameter tuning of the LSS-ALA algorithm to further improve convergence speed and solution quality for China UAV drone trajectory planning
- Implementing the proposed framework on embedded flight controllers for real-world field testing on actual bridge structures
The proposed methodology represents a significant step toward fully autonomous China UAV drone inspection of complex infrastructure, with potential applications extending to other challenging environments such as industrial plants, transmission towers, and offshore platforms.
8. Conclusion
This research presents a comprehensive framework for safe and adaptive trajectory planning of China UAV drone in complex bridge inspection scenarios, addressing the critical challenges of BIM data redundancy and susceptibility of optimization algorithms to local optima. The key findings and contributions of our work are summarized as follows:
First, we developed a lightweight environment modeling approach based on semantic filtering and component surface normal vector outward extrapolation. By extracting key bridge component features and delineating safe observation zones, we effectively transformed the high-dimensional physical space into low-dimensional static boundary constraints, significantly reducing the computational burden of path solving for China UAV drone.
Second, we constructed the LSS-ALA algorithm to overcome the stagnation tendency of the standard ALA. The LHS strategy optimizes population initialization to improve solution space distribution uniformity, while the embedded sparrow search vigilance mechanism reconstructs position update rules to effectively balance global exploration and local exploitation, overcoming premature convergence issues that commonly affect China UAV drone trajectory optimization in dense obstacle environments.
Third, we established a multi-constraint trajectory planning framework that unifies flight distance, onboard energy consumption, attitude yaw angle, and obstacle avoidance safety margins into a comprehensive objective function. The cubic B-spline interpolation technique fits discrete waypoints into a continuous smooth flight trajectory satisfying the dynamic constraints of China UAV drone. Comparative experiments on the CEC2022 benchmark suite and a realistic long-span cable-stayed bridge model demonstrate that LSS-ALA achieves significant advantages over GWO, SABO, and other algorithms in terms of optimization accuracy and robustness. The generated trajectories satisfy the safety and kinematic constraints for China UAV drone operations in complex environments, with a 17.2% reduction in shortest path length compared to GWO and a 27.6% reduction in cumulative turning angle compared to SABO, while maintaining a low standard deviation of 21.36 across multiple independent runs.
The proposed methodology provides a solid foundation for fully autonomous China UAV drone bridge inspection, with potential for extension to other infrastructure monitoring applications requiring safe navigation through complex geometric environments.