Survey on Comprehensive Visual Perception Technology for Future Air–Ground Intelligent Transportation Vehicles in All Scenarios

Guopeng Ren, Fei Chen, Shichun Yang, Fan Zhou, Bin Xu2025 · DOI

TL;DR

This survey paper explores the critical role of visual perception technology for future air-ground intelligent transportation vehicles, particularly in the burgeoning low-altitude economy. It synthesizes current research across five key areas—visual SLAM, Bird's-Eye View (BEV) perception, image enhancement, cloud computing optimization, and edge deployment—to highlight challenges and future directions for robust, all-scenario perception.

What Problem Does This Paper Address?

The rapid development of the "low-altitude economy," encompassing flying cars and other aerial vehicles, presents a significant challenge for current perception systems. Existing 2D autonomous driving perception methods are insufficient for the complex 3D environments these vehicles will navigate. Specifically, challenges include:

  • Global High-Precision Mapping: Creating detailed 3D maps of low-altitude environments.
  • Local Environmental Understanding: Accurately identifying and tracking traffic participants and other objects in real-time.
  • Extreme Condition Perception: Extracting reliable visual information under adverse conditions like poor lighting or dynamic scenes.

Addressing these gaps is crucial for ensuring the safety, efficiency, and widespread adoption of air-ground intelligent transportation.

Methods

The paper surveys existing literature and categorizes current approaches into five main areas:

  • Visual SLAM (Simultaneous Localization and Mapping): Focuses on using visual sensors for both positioning and mapping. It discusses feature-based and direct SLAM methods, highlighting advancements like the Smooth Variable Structure Filter (SVSF) for improved robustness in dynamic and non-textured environments.
  • Bird's-Eye View (BEV) Perception: Explores techniques to transform perspective camera views into a top-down BEV representation, which is more intuitive for tasks like navigation and object detection. Methods discussed include traditional Inverse Perspective Mapping (IPM) and newer deep learning approaches (depth-centric, MLP-centric, transformer-centric).
  • Image Enhancement: Reviews techniques to improve image quality for better perception. This includes histogram equalization, wavelet transforms, partial differential equations, and the Retinex algorithm, all aimed at enhancing contrast, reducing noise, and improving detail visibility.
  • Cloud Computing for Optimization: Discusses leveraging cloud resources to offload computationally intensive perception tasks, thereby extending the endurance of low-altitude vehicles. It covers cloud service models (IaaS, PaaS, SaaS) and optimization algorithms like ant colony optimization and particle swarm optimization.
  • Edge Deployment for Rapid Sensing: Examines strategies for deploying perception algorithms on edge nodes for low-latency processing and rapid response. This includes centralized and decentralized approaches, with a focus on task offloading techniques using machine learning and federated learning.

Datasets

The paper references several datasets used in the field, including:

  • CVT Dataset: Provides ground images with geographical markers, photographed pictures with timestamps, and synchronized high-altitude images across different seasons and times of day.
  • NPS-Drones: A collection of US Navy drone videos captured with GoPro-3 cameras.
  • FL-Drones: Photos of drones in various indoor and outdoor settings, with varying drone shapes and resolutions.
  • DTB70: A UAV tracking dataset featuring photos taken at high speeds, resulting in fuzzy images suitable for high-speed target recognition training.
  • UAV-Human Dataset: A diverse dataset captured by aerial drones in urban and rural settings, encompassing various conditions, subjects, and activities for motion recognition and posture estimation.
  • UAV-Trimar Dataset: A multi-modal dataset including RGB, skeleton, depth, IR, fisheye, and night-vision data, along with motion recognition targets and frames for posture estimation, re-identification, and attribute recognition.

Key Insights and Ideas

  • The "Low-Altitude Economy" is a Game Changer: The paper emphasizes that the rise of aerial vehicles necessitates a paradigm shift in perception technology beyond current automotive standards.
  • Visual SLAM is Key for Low-Altitude Navigation: Visual SLAM offers advantages over LiDAR for flying vehicles due to its ability to capture rich texture and provide high frame rates, crucial for accurate positioning and control.
  • BEV Perception is Essential for Intuitive Understanding: Transforming perspective views to BEV is vital for tasks like navigation planning and object detection, offering a more holistic environmental representation.
  • Image Enhancement is Not Just About Aesthetics: Techniques like histogram equalization and wavelet transforms are critical for extracting meaningful information from challenging visual inputs (e.g., low light, fog).
  • Hybrid Cloud-Edge Architectures are the Future: For real-time perception in resource-constrained aerial vehicles, a combination of cloud computing for heavy processing and edge computing for immediate responses is the most promising approach.

Key Results

The paper is a survey, so it doesn't present new experimental results. Instead, it synthesizes findings from existing research:

  • SVSF-based SLAM algorithms demonstrate superior performance and stability compared to conventional filter-based approaches, particularly in environments with noise and dynamic elements.
  • Transformer-based methods for BEV perception show strong relationship modeling capabilities and achieve good performance, though challenges remain in handling severe distortions.
  • Various image enhancement techniques (HE, wavelet, PDE, Retinex) offer improvements in contrast and detail but often face limitations in adaptability to complex, dynamic scenes or may sacrifice image detail.
  • Cloud computing algorithms (e.g., ant colony optimization, PSO, DGWO) can significantly optimize resource allocation and reduce task execution times.
  • Edge computing strategies using federated learning and actor-critic approaches show promise in reducing latency and energy usage for offloading tasks.

Limitations and Future Work

The authors acknowledge several limitations in current research and point towards future directions:

  • Real-time Performance: Many algorithms still struggle to meet the stringent real-time requirements for dynamic aerial navigation.
  • Adaptability to Dynamic Environments: Handling constantly changing scenes with moving objects remains a significant challenge.
  • Feature Extraction in Non-Textured Environments: Visual SLAM methods can be hindered in areas lacking distinct features.
  • Multi-Sensor Data Fusion: Integrating data from various sensors (cameras, LiDAR, IMU) effectively is complex.
  • Hardware Resource Optimization: Efficiently utilizing limited onboard computational resources is crucial.
  • Robustness and Generalization: Algorithms need to be more robust to varying illumination, weather, and complex scenarios.

Future work should focus on developing more accurate, robust, and real-time perception algorithms that can generalize across diverse scenarios, improve data fusion capabilities, and enable rapid deployment on edge devices. The integration of advanced deep learning techniques, coupled with efficient cloud and edge computing strategies, will be key to realizing the full potential of air-ground intelligent transportation.

Summary updated

Abstract

As an essential part of the low-altitude economy, low-altitude carriers are an important cornerstone of its development and a new industry that cannot be ignored strategically. However, it is difficult for the existing two-dimensional vehicle autonomous driving perception scheme to meet the needs of general key technologies for all-scene perception such as the global high-precision map construction of low-altitude vehicles in a three-dimensional space, the perception identification of local environmental traffic participants, and the extraction of key visual information under extreme conditions. Therefore, it is urgent to explore the development and verification of all-scene universal sensing technology for low-altitude intelligent vehicles. In this paper, the literature on vision-based urban rail transit and general perception technology in low-altitude flight environment is studied, and the paper summarizes the research status and innovation points from five aspects, namely the environment perception algorithm based on visual SLAM, the environment perception algorithm based on BEV, the environment perception algorithm based on image enhancement, the performance optimization of the perception algorithm using cloud computing, and the rapid deployment of the perception algorithm using edge nodes, and puts forward the future optimization direction of this topic.
BibTeX
@misc{ren2025survey,
  author = {Guopeng Ren and Fei Chen and Shichun Yang and Fan Zhou and Bin Xu},
  title = {{Survey on Comprehensive Visual Perception Technology for Future Air–Ground Intelligent Transportation Vehicles in All Scenarios}},
  year = {2025},
  doi = {10.3390/engproc2024080050},
}
Figure