With the rapid deployment of high-level autonomous driving technologies, intelligent connected vehicles (ICVs) face critical challenges in complex urban traffic environments. Limited single-vehicle perception range, insufficient coverage of blind areas, weak dynamic risk prediction capability, and poor path decision robustness in extreme scenarios severely impede the large-scale adoption of L4 and above autonomous driving. To address these issues, we propose a novel collaborative perception and path decision-making method leveraging UAV drones as mobile aerial sensing nodes. This method integrates unified environment modeling, a multi-objective optimization model, and a two-layer collaborative algorithm framework to achieve closed-loop optimization of perception and decision-making. Extensive simulations in medium-density and high-density urban traffic scenarios validate the effectiveness of our approach. In single-vehicle single-UAV drone collaborative scenarios, our improved Jump Point Search (JPS) algorithm reduces path decision time by up to 28.32% compared to the traditional A* algorithm, increases global perception coverage by 19.64%, and decreases comprehensive travel cost by up to 6.15%. Our research effectively compensates for the spatial limitations of single-vehicle perception in ICVs and provides theoretical and technical support for the implementation of air-ground integrated autonomous driving, offering significant engineering value for the deep integration of low-altitude economy and intelligent transportation systems.
1. Introduction
The fusion of vehicle intelligence and networking has made ICVs a core component of intelligent transportation systems. However, urban environments introduce severe perception bottlenecks: large vehicles, building occlusions, and adverse weather conditions create extended blind spots, delay dynamic obstacle prediction, and cause insufficient response to sudden traffic events. These issues become major barriers to commercializing L4 and above autonomous driving. UAV drones, with their high mobility, global view, and freedom from ground traffic constraints, offer a transformative solution. Acting as airborne mobile sensing nodes, they can build an air-ground collaborative perception system with ground ICVs, providing beyond-line-of-sight and full-area environmental information. This collaboration enhances both safety and efficiency of path decision-making. Recent regulatory frameworks such as the “Interim Regulations on the Flight Management of Unmanned Aerial Vehicles” and the large-scale promotion of “Vehicle-Road-Cloud Integration” technologies have provided policy and technical support for the collaborative application of UAV drones and ICVs. However, efficient collaborative perception task scheduling and integrated path decision-making remain the core challenges for large-scale deployment of air-ground collaborative autonomous driving.
Existing research has addressed aspects of this problem. Some studies have developed UAV drone-ICV collaborative perception frameworks using visual fusion to improve target detection accuracy in complex scenes. Others optimized UAV drone node placement to achieve full perception coverage of urban road blind spots. In path decision-making, improved A* algorithms have been applied for trajectory optimization, and JPS algorithms have been enhanced to reduce node expansions and improve planning efficiency in dynamic environments. Multi-agent reinforcement learning, particularly Multi-Agent Proximal Policy Optimization (MAPPO), has also gained attention for collaborative task allocation in vehicle networks due to its stable training and suitability for multi-agent scenarios. Despite these advances, two key gaps remain: first, most methods treat perception and path decision as independent problems, lacking tight coupling necessary for closed-loop “perception-decision-execution” optimization; second, existing approaches struggle to simultaneously handle real-time perception scheduling and integrated path planning in multi-vehicle multi-UAV drone dynamic traffic environments, resulting in insufficient robustness. Our work directly addresses these gaps by presenting a unified two-layer collaborative algorithm that integrates the improved JPS algorithm with MAPPO, validated through comprehensive simulations.
2. System Modeling and Problem Formulation
2.1 Core Operational Modes of the Vehicle-UAV Drone Collaborative System
Our collaborative system consists of multi-rotor UAV drones and ground ICVs operating in an air-ground integrated architecture. The UAV drones communicate wirelessly with ICVs to exchange real-time perception information and coordinate decisions. We define two typical modes of operation:
- Accompanying Collaboration: A UAV drone follows the host ICV synchronously, flying within a predefined airspace above the vehicle. It continuously collects full surrounding environmental information, compensating for blind spots in the front, rear, and sides, thus supporting lane keeping, lane changing, and obstacle avoidance.
- Forward Collaboration: A UAV drone pre-flies to the upcoming road segments along the vehicle’s planned path, gathering beyond-visual-range traffic information such as intersection blind spots, accidents, and congestion. This enables path re-planning and local decision adjustment in advance.
In compliance with the “Interim Regulations on the Flight Management of Unmanned Aerial Vehicles” and urban airspace management norms, we divide the low-altitude airspace from ground to 120 m into three vertical flight layers. The altitude range from 30 m to 80 m is designated as the dedicated collaborative sensing airspace for UAV drones. Horizontally, the airspace is partitioned into grid management units corresponding to the urban road network, forming a “vertical layering + road network zoning” airspace management system to ensure spatial decoupling and safety isolation between UAV drone flights and ground traffic.
2.2 Air-Ground Integrated Discretized Modeling Using Grid Method
We adopt a grid-based method to discretize both the urban ground traffic environment and the low-altitude sensing airspace into a unified 3D grid map. This provides a standardized environmental input for collaborative perception and path decision-making. Let the 2D planar range of the study area be X∈[0, Xmax], Y∈[0, Ymax], and the vertical height range be Z∈[0, Zmax]. The entire region is partitioned into Nx × Ny × Nz cubic grids with side length L:
$$
N_x = \lceil X_{max}/L \rceil, \quad N_y = \lceil Y_{max}/L \rceil, \quad N_z = \lceil Z_{max}/L \rceil
$$
Each grid is assigned a state value S(x,y,z) to represent its passability and perception attributes:
- S=0: passable/perceivable area
- S=1: ground obstacle/building area
- S=2: no-fly zone/traffic control area
- S=3: perception blind area
ICVs are allowed to move only in grids with S=0 at ground level (Z=0). UAV drones are permitted to fly only in grids with S=0. This digitization captures ground traffic constraints, airspace constraints, obstacles, and blind spots in a unified manner.
2.3 Comprehensive Risk Map Construction for Vehicle-UAV Drone Collaboration
We classify risks in the collaborative system into three categories: ground traffic risk, aerial flight risk, and perception blind area risk. These are integrated through multi-layer weighted superposition to create a comprehensive risk map, providing detailed safety constraint information for path decision-making. Our risk map consists of five core layers:
- Traffic Flow Density Layer: Quantifies the impact of ground congestion on vehicle passage and UAV drone perception. The traffic flow density risk at point (x,y) is calculated as:
$$
R_{tra}(x,y) = \sum_{i=1}^{n} \frac{p_i}{1 + d_i}
$$
where p_i is the traffic density at the i-th grid cell, and d_i is the distance from the target point to the center of that grid cell. - Traffic Risk Point Layer: High-risk areas such as intersections, accident black spots, schools, and hospitals are assigned weights based on their risk levels and adjusted by time factors:
$$
R_{risk}(x,y) = \sum_{k} \frac{\omega_k \mu_t}{1 + d_k}
$$
where ω_k is the area-type weight (1 for high, 0.5 for medium, 0.2 for low risk), μ_t is the time factor (1.0 for peak, 0.6 for off-peak, 0.3 for night), and d_k is the distance from the target point to the boundary of the risk area. - Obstacle and No-Fly Zone Layer: Penalizes proximity to obstacles and no-fly zones using attenuation functions:
$$
R_{obs}(x,y) =
\begin{cases}
\lambda \cdot e^{-d_{obs}} + \lambda_f \cdot e^{-d_f}, & d_{obs} > 0 \\
1, & \text{in obstacle}
\end{cases}
$$
where d_{obs} is the distance to the nearest obstacle, d_f is the distance to the nearest no-fly zone boundary, and λ, λ_f are decay coefficients between 0.1 and 1.0. - Perception Blind Area Risk Layer: Quantifies the risk due to single-vehicle perception blind spots:
$$
R_{blind}(x,y) = \omega_b \cdot \frac{S_{blind}}{S_{total}}
$$
where S_{blind} is the blind area size around the target point, S_{total} is the total perceivable area, and ω_b is the blind area risk weight. - Aerial Operation Risk Layer: Models collision risk and signal interference risk for low-altitude UAV drone flights:
$$
R_{air}(x,y,z) = \omega_{air1} e^{-d_{air}} + \omega_{air2} |z – z_0|
$$
where d_{air} is the distance to other flying units, z_0 is the optimal sensing altitude, and ω_{air1}, ω_{air2} are risk weights.
The five risk layers are normalized and combined using weighted summation:
$$
R_{total}(x,y,z) = \omega_1 R_{tra} + \omega_2 R_{risk} + \omega_3 R_{obs} + \omega_4 R_{blind} + \omega_5 R_{air}
$$
with weights ω_1 through ω_5 summing to 1. This comprehensive risk map serves as the safety constraint for collaborative perception and path decision-making.
2.4 UAV Drone Node Layout Optimization Using P-Median Model
UAV drone takeoff/landing and standby nodes are critical infrastructure. We use the P-median model to optimize node placement. The core objectives are to maximize perception coverage of the entire road network while minimizing the total weighted distance cost from vehicles to the nearest service node. The objective function is:
$$
\begin{aligned}
\min \quad & Z = \sum_{j \in J} f_j y_j + \sum_{i \in I} \sum_{j \in J} c_{ij} q_i d_{ij} x_{ij} \\
\max \quad & C_{cover} = \frac{\sum_{i \in I} S_{cover}(i)}{S_{total}} \\
\end{aligned}
$$
where I is the set of road network demand points, J is the set of candidate nodes, f_j is the fixed construction cost of candidate node j, c_{ij} is the unit-km transportation cost from node j to demand point i, d_{ij} is the flight distance between them, q_i is the traffic flow at demand point i, S_{cover}(i) is the perception coverage area at demand point i, and S_{total} is the total study area. Decision variables: y_j ∈ {0,1} (1 if node j is selected), x_{ij} ∈ {0,1} (1 if demand point i is served by node j). Constraints include:
$$
\begin{cases}
\sum_{j \in J} x_{ij} = 1, & \forall i \in I \\
x_{ij} \le y_j, & \forall i \in I, j \in J \\
\sum_{j \in J} y_j = P \\
\sum_{i \in I} q_i x_{ij} \le Q_{max}, & \forall j \in J \\
d_{ij} \le D_{max}, & \forall i \in I, j \in J \\
d_{jj’} \ge D_{min}, & \forall j, j’ \in J, j \ne j’
\end{cases}
$$
where P is the number of selected nodes, Q_{max} is the maximum UAV drone capacity per node, D_{max} is the maximum single-flight distance, and D_{min} is the minimum safe distance between nodes.
3. Multi-Objective Optimization Model for Collaborative Path Decision-Making
Given the vehicle-UAV drone collaborative scenario, we construct a multi-constraint multi-objective path decision optimization model with the primary goals of maximizing global perception coverage and minimizing comprehensive travel cost.
3.1 Objective Function Construction
Our total cost objective integrates perception coverage, travel efficiency, collaborative cost, and energy consumption:
$$
C_{total} = \omega_1 \cdot (-C_{cover}) + \omega_2 \cdot F_{vehicle} + \omega_3 \cdot F_{uav} + \omega_4 \cdot F_{cost}
$$
where ω_1 through ω_4 are weight coefficients for perception coverage, vehicle travel cost, UAV drone flight cost, and collaborative cost, respectively, satisfying ω_1+ω_2+ω_3+ω_4=1. The negative sign before C_{cover} reflects minimization orientation.
- Perception Coverage:
$$
C_{cover} = \frac{S_{vehicle} \cup S_{uav}}{S_{target}} \times 100\%
$$
where S_{vehicle} is the single-vehicle perception coverage area, S_{uav} is the UAV drone perception coverage area, and S_{target} is the total target segment area. - Vehicle Travel Cost:
$$
F_{vehicle} = \sum_{k=1}^{n-1} \left( \alpha_1 l_k + \alpha_2 t_k + \alpha_3 R_{road}(k) \right)
$$
where l_k is the path length between consecutive nodes, t_k is the travel time, R_{road}(k) is the ground risk value at node k, and α_1, α_2, α_3 are weights. - UAV Drone Flight Cost:
$$
F_{uav} = \sum_{k=1}^{m-1} \left( \beta_1 d_k + \beta_2 e_k + \beta_3 R_{air}(k) \right)
$$
where d_k is the flight distance, e_k is the energy consumption, R_{air}(k) is the air risk value, and β_1, β_2, β_3 are weights. - Collaborative Cost:
$$
F_{cost} = \gamma_1 \left| t_{vehicle} – t_{uav} \right| + \gamma_2 n_{task}
$$
where t_{vehicle} and t_{uav} are the arrival times of the vehicle and UAV drone at the target node, n_{task} is the number of task scheduling events, and γ_1, γ_2 are weights.
The weight coefficients are determined using the Analytic Hierarchy Process (AHP) with consistency check (CR=0.072 < 0.1). The baseline weights are: ω_1=0.4, ω_2=0.3, ω_3=0.2, ω_4=0.1, prioritizing perception safety, then efficiency, then UAV drone energy, then scheduling cost.
3.2 Constraints
ICV Constraints:
- Dynamics: v_k ≤ v_{max}, a_{min} ≤ a_k ≤ a_{max}, R_k ≥ R_{min}
- Traffic rules: no reverse, no lane departure, obey traffic signals.
- Collision avoidance: all grid states S=0 along the path; maintain safe distance from dynamic obstacles.
UAV Drone Constraints:
- Performance: d_k ≤ D_{max}, v_{uav} ≤ v_{uav,max}, |θ_k| ≤ θ_{max}, z_{min} ≤ z_k ≤ z_{max}, E_0 – ΣΔe_k ≥ E_{res}
- Airspace: only fly in dedicated collaborative airspace; avoid no-fly zones; maintain safe distance from other flying objects.
- Perception: flight path must cover target segments; horizontal distance to the vehicle ≤ D_{commax} for real-time data transmission.
Collaborative Constraints:
- Communication distance ≤ D_{commax}
- Time synchronization: |t_{vehicle,k} – t_{uav,k}| ≤ T_{max}
- Perception collaboration: UAV drone sensing must cover blind spots and high-risk areas along the vehicle’s path.
4. Two-Layer Collaborative Perception and Path Decision Algorithm
We design a nested two-layer collaborative optimization algorithm where the upper layer uses MAPPO for dynamic multi-UAV drone multi-vehicle perception task allocation, and the lower layer uses an improved JPS algorithm for fine-grained vehicle-UAV drone integrated path planning.
4.1 Improved JPS Algorithm
Traditional JPS suffers from rigidity in dynamic environments and insufficient adaptability to collaboration constraints. Our improvements include:
- Bidirectional Search: We simultaneously perform forward search from the start S and backward search from the goal E using separate open lists (open_start, open_end) and closed lists (close_start, close_end). In each iteration, the node with minimal cost from each list is expanded for jump points. When the two searches meet, the complete path is formed by concatenation. Cost functions:
$$
f_{start}(n) = g_{start}(n) + h_{start}(n), \quad
f_{end}(m) = g_{end}(m) + h_{end}(m)
$$
where g_{start}(n) is actual cost from start to n, h_{start}(n) is heuristic from n to end, etc. - Dynamic Target Following: For the vehicle search, the UAV drone’s real-time sensed optimal forward node is used as a temporary target. For the UAV drone search, the vehicle’s current and predicted positions serve as dynamic targets. After each search cycle, the temporary target is updated based on real-time states, maintaining flexibility.
- Heuristic Function Optimization: We incorporate global risk and perception coverage into the heuristic:
$$
h(n) = \mu_1 d(n, end) + \mu_2 R_{total}(n) + \mu_3 \frac{1}{C_{cover}(n)}
$$
where d(n,end) is Euclidean distance, R_{total}(n) is comprehensive risk, C_{cover}(n) is perception coverage, and μ_1, μ_2, μ_3 are weights. This balances safety, efficiency, and coverage. - Path Smoothing for Coordination: After initial path generation:
- Vehicle path smoothing: Apply B-spline curve fitting to remove sharp turns, satisfying vehicle dynamics constraints (minimum turning radius, acceleration limits).
- UAV drone accompanying path optimization: Relax JPS’s grid-only diagonal movement by optimizing the UAV drone trajectory along the vehicle’s smoothed path, keeping the UAV drone at optimal sensing positions while minimizing turns and energy consumption.
4.2 Upper-Layer MAPPO for Collaborative Perception Task Allocation
For multi-UAV drone multi-ICV scenarios, we design a MAPPO-based algorithm to dynamically allocate perception tasks in real time. The state space and action space are defined as:
State Space:
$$
s = [V_{state}, U_{state}, E_{env}, T_{task}]
$$
where V_{state} includes positions, speeds, remaining energy, perception ranges, and planned paths of all ICVs; U_{state} includes positions, speeds, remaining energy, perception ranges, and current tasks of all UAV drones; E_{env} is the full traffic environment including traffic density, obstacle distribution, risk points, and airspace constraints; T_{task} is the set of pending tasks with locations, priorities, time windows, and coverage requirements.
Action Space: A discrete set of matchings:
$$
A = \{ (U_j, V_i, T_k) \ | \ j \in [1,N_u], i \in [1,N_v], k \in [1,N_t] \}
$$
where each triple assigns the k-th perception task to the j-th UAV drone to serve the i-th ICV.
Reward Function: Combined immediate and terminal rewards:
$$
R = R_{cover} + R_{eff} – R_{cost} – R_{punish}
$$
where R_{cover} rewards increased perception coverage, R_{eff} rewards efficient travel, R_{cost} penalizes collaborative overhead, and R_{punish} penalizes violations (e.g., exceeding communication distance).
5. Simulation Experiments and Results
We conducted our simulation experiments using MATLAB/Simulink with the Prescan traffic simulation tool to build realistic urban scenarios. The ICV parameters were set based on typical passenger autonomous vehicles, and the UAV drone parameters were set based on commercial multi-rotor drones. Key parameters are shown in Table 1.
| Category | Parameter | Value |
|---|---|---|
| ICV | Max Speed | 60 km/h |
| Max Acceleration | 3 m/s² | |
| Max Deceleration | 8 m/s² | |
| Min Turning Radius | 5 m | |
| UAV Drone | Max Cruising Speed | 15 m/s |
| Max Flight Range | 30 km | |
| Max Flight Time | 55 min | |
| Sensing | ICV Perception Distance | 200 m |
| UAV Drone Perception Radius | 500 m | |
| Coordination | Max Communication Distance | 1000 m |
| Max Time Deviation | 2 s | |
| Optimal UAV Drone Altitude | 50 m |
Two typical urban scenarios were designed:
- Medium-Density Urban Scenario: 3 km × 3 km area with 4 main roads, 8 secondary roads, 12 intersections, medium traffic density, 30% obstacle coverage, 2 ICVs, 2 UAV drones.
- High-Density Urban Scenario: 5 km × 5 km area with 6 main roads, 15 secondary roads, 24 intersections, high traffic density, over 50% obstacle coverage, high-risk zones (schools, hospitals, shopping areas), 5 ICVs, 4 UAV drones.
In our experiments, the average comprehensive cost directly corresponds to the total cost objective defined in Section 3.1. To eliminate dimensional influences, each sub-cost was min-max normalized, then weighted using the baseline weights. Smaller values indicate better overall performance. For each scenario, we ran 20 independent simulations for our improved JPS algorithm and compared it with traditional A*, Whale Optimization Algorithm (WOA), Grey Wolf Optimizer (GWO), and Particle Swarm Optimization (PSO). The average results are reported in Tables 2 and 3.
| Algorithm | Vehicle Path Length (m) | Planning Time (s) | Perception Coverage | Comprehensive Cost |
|---|---|---|---|---|
| A* | 2864.32 | 2.12 | 72.35% | 142.68 |
| WOA | 2812.57 | 1.26 | 76.48% | 137.59 |
| GWO | 2798.43 | 1.15 | 77.62% | 135.84 |
| PSO | 2805.69 | 1.32 | 75.93% | 138.26 |
| Improved JPS | 2756.81 | 1.08 | 90.27% | 131.25 |
| Algorithm | Vehicle Path Length (m) | Planning Time (s) | Perception Coverage | Comprehensive Cost |
|---|---|---|---|---|
| A* | 5638.74 | 9.36 | 65.28% | 297.43 |
| WOA | 5726.31 | 6.92 | 68.74% | 291.56 |
| GWO | 5612.85 | 8.05 | 71.36% | 283.47 |
| PSO | 5703.49 | 8.24 | 69.52% | 289.62 |
| Improved JPS | 5482.36 | 6.71 | 84.92% | 279.14 |
From the tables, in the medium-density scenario, our improved JPS algorithm reduces vehicle path length by 3.75%, planning time by 49.06%, improves perception coverage by 24.77%, and reduces comprehensive cost by 8.01% compared to traditional A*. Against swarm intelligence algorithms (WOA, GWO, PSO), it maintains clear advantages in all metrics. In the high-density scenario, the improvements are even more pronounced: path length reduced by 2.77% vs A*, planning time reduced by 28.32%, perception coverage increased by 19.64%, and comprehensive cost reduced by 6.15%. Compared to other algorithms, path length is reduced by up to 4.26%, perception coverage increased by up to 16.16%, and comprehensive cost reduced by up to 4.33%. These results validate the robustness and superiority of our algorithm in complex urban environments.
To analyze the effect of weight coefficients, we conducted sensitivity experiments with four different weight combinations in the high-density scenario, repeating 20 times. The results are shown in Table 4.
| ID | ω₁ | ω₂ | ω₃ | ω₄ | Avg Path Length (m) | Avg Perception Coverage | Avg Comprehensive Cost |
|---|---|---|---|---|---|---|---|
| A | 0.5 | 0.3 | 0.1 | 0.1 | 5526.47 | 86.35% | 282.64 |
| B (baseline) | 0.4 | 0.3 | 0.2 | 0.1 | 5482.36 | 84.92% | 279.14 |
| C | 0.3 | 0.4 | 0.2 | 0.1 | 5457.82 | 78.64% | 283.75 |
| D | 0.2 | 0.5 | 0.2 | 0.1 | 5432.19 | 72.38% | 288.46 |
The trend shows that as the perception coverage weight ω₁ decreases and the vehicle travel weight ω₂ increases, path length decreases but perception coverage drops sharply, and comprehensive cost first decreases then increases. The baseline combination B (ω₁=0.4, ω₂=0.3, ω₃=0.2, ω₄=0.1) achieves the best balance, yielding the lowest comprehensive cost (279.14). This provides a reference for weight selection in urban vehicle-UAV drone collaborative scenarios.
6. Conclusion
Through extensive multi-scenario simulations, we draw the following conclusions. The comprehensive risk map incorporating traffic density, risk points, airspace constraints, blind spots, and obstacles effectively quantifies operational risks in complex urban environments, providing fine-grained safety constraints for vehicle-UAV drone collaborative perception and path decision-making. Our UAV drone node layout optimization using the P-median model increases global perception coverage by 17.32% and reduces average service distance by 10.47% compared to uniform placement, enabling efficient infrastructure planning. The improved JPS algorithm, with bidirectional search, dynamic target following, heuristic optimization, and path smoothing, significantly enhances the efficiency and performance of vehicle-UAV drone integrated path planning. In the high-density urban scenario, it reduces path decision time by 28.32% over traditional A*, increases perception coverage by 19.64%, and decreases comprehensive travel cost by 6.15%. Our approach effectively addresses the core pain points of limited single-vehicle perception, blind spot coverage, and weak dynamic risk prediction, achieving closed-loop optimization of collaborative perception and path decision-making. Future work will focus on extending the method to extreme weather and sudden traffic event scenarios, reducing the computational burden of the centralized MAPPO training, and developing adaptive weight adjustment mechanisms to further improve robustness and practical deployment potential.