Our research team has systematically investigated the low-altitude emergency unmanned aerial vehicle (UAV) control strategies for mountainous highway tunnel groups, a critical yet underexplored domain in intelligent transportation systems. This study focuses on a representative spiral tunnel group located in a deep mountainous region, characterized by enclosed environments, complex terrains, and significant elevation changes. The primary objective is to develop a precise control framework that aligns UAV drones’ patrol frequency with the real-time driving risk levels, thereby enhancing emergency response efficiency and operational safety. Through real-vehicle experiments, multi-dimensional data encompassing driver, vehicle, road, and environmental factors were collected. A driving risk quantification model was constructed using factor analysis, and four risk levels were identified: low, medium-low, medium-high, and high. By introducing the concept of risk exposure, we established a differentiated inspection frequency calculation method for UAV drones. The resulting control strategy provides a quantitative basis for deploying UAV drones in tunnel groups, ensuring that high-risk segments receive more frequent surveillance. This work contributes to the theoretical foundation of low-altitude emergency UAV management and offers practical guidelines for improving traffic safety in complex tunnel systems.
The rapid expansion of expressway networks into mountainous regions has led to an increasing number of tunnel groups. These tunnels often feature spiral alignments, steep gradients, and narrow cross-sections, creating unique challenges for traffic safety and emergency response. Traditional ground-based rescue methods suffer from delayed information acquisition and poor coordination. In contrast, UAV drones offer high mobility, quick deployment, and the ability to carry diverse sensors, making them ideal for reconnaissance, communication relay, and material delivery. However, the confined space, GPS signal blockage, and complex airflow inside tunnels impose severe constraints on UAV drones’ flight stability and positioning accuracy. Moreover, the lack of tailored control strategies for such special environments hinders their effective application. Therefore, our research aims to bridge this gap by proposing a risk-based UAV control strategy specifically designed for mountainous highway tunnel groups.
Real-Vehicle Driving Experiment
To obtain realistic and high-quality data, we conducted a real-vehicle driving experiment on a typical spiral tunnel group in a mountainous area. The experiment route consisted of three consecutive spiral tunnels with a total length of approximately 5.6 km. The tunnels have varying radii ranging from 730 m to 2500 m, gradients around 2.45%, and design speeds of 80 km/h. Forty participants (30 males, 10 females) aged 20 to 70 years (mean 41.5, SD 15.6) were recruited through stratified sampling based on age, gender, occupation, and education level. All participants signed informed consent forms and completed pre-experiment questionnaires. To minimize individual differences, data standardization was applied during processing, and the experiment procedures were strictly controlled.
The experimental equipment comprised four categories: vehicle, data collection, recording, and output devices. The vehicle was an automatic transmission car. Collection devices included an On-Board Diagnostics (OBD) system for vehicle speed and acceleration, an Inertial Measurement Unit (IMU) for lateral offset and steering, a luxmeter for illuminance, and a noise meter. Driver physiological data were captured using Tobii Glasses 3 eye-tracker and ErgoLAB wearable physiological recorder, measuring heart rate and pupil area. Two cameras recorded the driver’s behavior and road conditions. All data were synchronized using ErgoLAB multi-channel module.
The experiment route started from Wanxianshan Toll Station to Xiyagou Toll Station, covering 38.5 km in total. Round trips were conducted at different times of the day. In the preparation phase, participants completed questionnaires, were briefed on safety requirements, and underwent equipment calibration. During the formal experiment, all devices recorded simultaneously while drivers operated normally. Data completeness was verified after each run. A summary of the experimental parameters is presented in Table 1.
| Parameter | Value / Description |
|---|---|
| Total tunnel length (3 sections) | 5,630 m (1,330 m + 2,100 m + 2,200 m) |
| Minimum curvature radius | 730 m |
| Maximum gradient | 2.45% |
| Design speed | 80 km/h |
| Number of participants | 40 (30 male, 10 female) |
| Age range (mean ± SD) | 20–70 (41.5 ± 15.6 years) |
| Measured variables | Speed, longitudinal acceleration, lateral offset, steering angle, heart rate, pupil area, road curvature, road gradient, illuminance |
| Data collection devices | OBD, IMU, luxmeter, noise meter, Tobii Glasses 3, ErgoLAB recorder |
| Experiment duration per run | Approximately 30–40 minutes |
Driving Risk Quantification
From the raw experimental data, we extracted nine candidate indicators covering the “driver-vehicle-road-environment” system: heart rate (HR), pupil area (PA), vehicle speed (v), longitudinal acceleration (a), lateral offset (d), steering wheel angle (SW), illuminance (Lx), road radius (R), and road gradient. A Pearson correlation analysis revealed significant correlations (p < 0.001) among all variables except road gradient. Since road gradient showed no significant correlation with others, it was excluded from further analysis. The remaining eight variables exhibited moderate to strong intercorrelations, making them suitable for factor analysis.
The Kaiser-Meyer-Olkin (KMO) measure of sampling adequacy was 0.726, indicating a sufficient level of shared variance. Bartlett’s test of sphericity yielded a Chi-square statistic with p < 0.001, confirming that the correlation matrix was not an identity matrix. Therefore, factor analysis was appropriate. Principal component extraction with varimax rotation produced three factors with eigenvalues greater than 1.0, explaining 86.06% of the total variance. Table 2 shows the total variance explained.
| Component | Initial Eigenvalue | % of Variance | Cumulative % |
|---|---|---|---|
| Factor 1 (Vehicle Lateral Stability) | 3.872 | 48.40 | 48.40 |
| Factor 2 (Physiological Load) | 2.045 | 25.56 | 73.96 |
| Factor 3 (Vehicle Longitudinal Control) | 1.028 | 12.10 | 86.06 |
| Factor 4 | 0.466 | 5.82 | 91.88 |
| Factor 5 | 0.350 | 4.36 | 96.24 |
| Factor 6 | 0.173 | 2.16 | 98.41 |
| Factor 7 | 0.098 | 1.23 | 99.64 |
| Factor 8 | 0.029 | 0.36 | 100.00 |
After rotation, the factor loadings and score coefficient matrices were obtained. Factor 1 (Vehicle Lateral Stability) had high loadings from lateral offset (0.831), steering wheel angle (-0.959), and road radius (-0.958). Factor 2 (Physiological Load) was dominated by heart rate (0.460), pupil area (-0.951), and illuminance (0.927). Factor 3 (Vehicle Longitudinal Control) was primarily associated with speed (0.893) and acceleration (0.967). The factor score functions are given as linear combinations of standardized variables:
$$
\begin{aligned}
SF_1 &= -0.151 z_{\text{HR}} – 0.077 z_{\text{PA}} + 0.007 z_v + 0.016 z_a + 0.220 z_d – 0.262 z_{\text{SW}} + 0.010 z_{\text{Lx}} – 0.256 z_R \\
SF_2 &= 0.238 z_{\text{HR}} – 0.511 z_{\text{PA}} + 0.040 z_v + 0.073 z_a + 0.003 z_d – 0.093 z_{\text{SW}} + 0.448 z_{\text{Lx}} – 0.035 z_R \\
SF_3 &= 0.224 z_{\text{HR}} – 0.140 z_{\text{PA}} + 0.239 z_v + 0.916 z_a – 0.175 z_d – 0.115 z_{\text{SW}} – 0.089 z_{\text{Lx}} – 0.122 z_R
\end{aligned}
$$
where each \(z\) denotes the standardized value (mean = 0, SD = 1). The composite driving risk score \(SF\) was computed by weighting the three factors according to their variance contributions:
$$
SF = \frac{1}{0.8606} \left(0.4741 SF_1 + 0.2513 SF_2 + 0.1352 SF_3\right)
$$
We then applied the k-means++ clustering algorithm on the SF scores to discretize driving risk into four levels. The thresholds are listed in Table 3. The risk distribution along the tunnel length was visualized, revealing clear spatial heterogeneity: the highest risks concentrated near the smallest curve radii (730 m), while lower risks occurred in wider sections.
| Risk Level | SF Score Range |
|---|---|
| Low | [1.25, 2.61) |
| Medium-Low | [2.61, 4.13) |
| Medium-High | [4.13, 5.50) |
| High | [5.50, 7.63] |
The proportion of each risk level across the three tunnel sections is summarized in Table 4. It is evident that the medium-low risk segment occupied the largest share (about 46–51%), while high-risk segments accounted for about 11–13%.
| Tunnel Section | Low | Medium-Low | Medium-High | High |
|---|---|---|---|---|
| Tunnel A | 15.06 | 51.13 | 22.54 | 11.27 |
| Tunnel B | 10.11 | 46.13 | 32.51 | 11.25 |
| Tunnel C | 5.87 | 46.33 | 35.09 | 12.71 |
UAV Drone Patrol Control Strategy
To translate the quantified driving risk into actionable UAV drone patrol schemes, we introduced the concept of risk exposure, borrowed from reliability engineering and risk-based inspection (RBI) methodology. The core idea is that the accumulated risk over a single inspection period should not exceed a predetermined acceptable threshold. For a given risk level \(g\), the risk exposure within one patrol cycle \(T_g\) is defined as:
$$
E_g = F_g^* \cdot T_g
$$
where \(F_g^*\) is the representative risk value for level \(g\), taken as the mid-point of its range. The four mid-points are:
$$
F_L^* = 1.93,\quad F_{ML}^* = 3.37,\quad F_{MH}^* = 4.81,\quad F_H^* = 6.56
$$
For each tunnel, a weighted average baseline risk value \(\bar{F}_w\) is computed using the proportion of each risk level from Table 4:
$$
\bar{F}_w = w_L F_L^* + w_{ML} F_{ML}^* + w_{MH} F_{MH}^* + w_H F_H^*
$$
The resulting baseline values for the three tunnels are: \(\bar{F}_{w,A}=3.84\), \(\bar{F}_{w,B}=4.05\), \(\bar{F}_{w,C}=4.20\). Assuming a standard patrol cycle \(T_{\text{baseline}} = 15\) minutes for the average risk, the risk exposure constant \(C\) for each tunnel is calibrated as:
$$
C = \bar{F}_w \cdot T_{\text{baseline}}
$$
Thus, the permissible risk exposure constants are: \(C_A = 57.60\), \(C_B = 60.75\), \(C_C = 63.00\). For any risk level \(g\) in tunnel \(k\), the required patrol cycle is obtained by setting \(E_g = C_k\):
$$
T_g^k = \frac{C_k}{F_g^*}
$$
The patrol frequency (in cycles per hour) is then:
$$
f_g^k = \frac{60}{T_g^k}
$$
Table 5 presents the computed patrol cycles and corresponding frequencies for each risk level across the three tunnels.
| Risk Level | Tunnel A | Tunnel B | Tunnel C | |||
|---|---|---|---|---|---|---|
| Cycle (min) | Freq (per h) | Cycle (min) | Freq (per h) | Cycle (min) | Freq (per h) | |
| Low | 29.84 | 2.01 | 31.48 | 1.91 | 32.64 | 1.84 |
| Medium-Low | 17.09 | 3.51 | 18.03 | 3.33 | 18.69 | 3.21 |
| Medium-High | 11.98 | 5.01 | 12.63 | 4.75 | 13.10 | 4.58 |
| High | 8.78 | 6.83 | 9.26 | 6.48 | 9.60 | 6.25 |
Based on the results, we formulate the following graded UAV drone control strategy:
- Low-risk segments: Patrol frequency ≈ 2 trips per hour (low-frequency inspection).
- Medium-low risk segments: Patrol frequency ≈ 3–4 trips per hour (routine inspection).
- Medium-high risk segments: Patrol frequency ≈ 5 trips per hour (enhanced inspection).
- High-risk segments: Patrol frequency ≥ 6 trips per hour (intensive inspection).
This strategy ensures that UAV drones allocate more resources to sections with elevated driving risk, thereby reducing the probability of undetected incidents. The approach is dynamic: if the risk profile changes (e.g., due to weather or traffic conditions), the patrol frequencies can be recalculated accordingly using the same framework. Moreover, the strategy can be integrated with a centralized UAV drone traffic management system to coordinate multiple drones, avoid collisions, and optimize flight paths in the confined airspace above the tunnel group.
Conclusion and Discussion
This study proposed a comprehensive low-altitude emergency UAV drone control strategy tailored for mountainous highway tunnel groups. By conducting real-vehicle experiments in a representative spiral tunnel group, we collected multi-source data and developed a driving risk quantification model using factor analysis. The model successfully reduced eight correlated indicators into three interpretable factors—vehicle lateral stability, physiological load, and vehicle longitudinal control—and produced a continuous risk score that was discretized into four meaningful levels via k-means++. The spatial visualization of risk revealed that the highest risk segments are predominantly located at sharp curve sections with small radii, aligning with intuitive expectations.
The main innovation lies in the integration of risk exposure theory from reliability engineering into UAV drone patrol scheduling. The risk exposure constant was calibrated using the weighted average risk of each tunnel and a baseline patrol cycle (15 minutes). Then, the required patrol cycle for each risk level was inversely derived, yielding differentiated frequencies ranging from 2 to over 6 patrols per hour. This quantitative link between risk and inspection intensity is a significant advancement over traditional qualitative or uniform patrol schemes.
Several limitations should be acknowledged. First, the experiment was conducted on a single tunnel group under normal driving conditions; the model may not fully capture extreme weather, traffic congestion, or accident scenarios. Future work should incorporate long-term data collection across multiple tunnel groups and consider dynamic factors such as real-time traffic density and meteorological conditions. Second, the baseline patrol cycle of 15 minutes was assumed based on typical highway emergency response times; its suitability should be validated through field trials with actual UAV drones. Third, the risk exposure constant was derived from the average risk of each tunnel, but a more refined approach could assign different constants to different segments based on consequence severity (e.g., proximity to tunnel portals, presence of emergency exits).
Despite these limitations, the proposed methodology provides a solid theoretical foundation for deploying UAV drones in emergency management of tunnel groups. The graded patrol strategy can be directly used to generate drone flight schedules, crew shift plans, and priority allocation of battery recharging stations. It also offers a template for regulators to design airspace management rules for low-altitude UAV drones in complex linear infrastructure. Our future research will focus on integrating real-time risk monitoring using onboard sensors and developing multi-UAV drone cooperative control algorithms that respect the calculated patrol frequencies while avoiding conflicts. The ultimate goal is to establish a fully autonomous, risk-aware UAV drone system that enhances the safety and resilience of mountainous highway tunnel operations.