Review on Vehicle Detection Technology for Unmanned Ground Vehicles

Qi Liu, Zirui Li, Shihua Yuan, Yuzheng Zhu, Xueyuan LiSensors 2021 · DOI

TL;DR

This paper provides a comprehensive review of vehicle detection technologies for Unmanned Ground Vehicles (UGVs), covering sensors, algorithms, simulation platforms, and datasets. It aims to offer a structured overview for researchers and engineers working on UGV perception systems, highlighting current trends and future research directions.

What Problem Does This Paper Address?

The increasing deployment of Unmanned Ground Vehicles (UGVs) in both civilian (e.g., autonomous driving) and military applications necessitates robust environmental perception capabilities. A critical component of this perception is accurate and reliable vehicle detection, which is fundamental for safe navigation, decision-making, and mission execution. This paper addresses the challenge of comprehensively reviewing the diverse and rapidly evolving landscape of vehicle detection technologies for UGVs, providing a structured guide to the state-of-the-art.

Methods

The paper categorizes vehicle detection methods based on the sensors used and the algorithmic approaches.

Sensor-Based Categories: * Vision-Based Methods: Discusses techniques using monocular cameras, stereo cameras, omni-direction cameras, event cameras, and infrared cameras. * Lidar-Based Methods: Covers approaches utilizing Lidar, including feature extraction, projection methods (spherical, front-view, bird-eye), voxel methods, and point-nets methods. * Radar-Based Methods: Reviews techniques employing millimeter-wave radar, such as registration, learning-based, end-to-end, and advanced imaging methods. * Infrared-Based Methods: Explores the use of infrared cameras, often complementing traditional vision methods for low-light conditions. * Event-Based Methods: Briefly touches upon the potential of event cameras for dynamic object detection. * Sensor-Fusion Methods: Examines how combining data from multiple sensors (e.g., Radar-Vision, Lidar-Vision) enhances robustness and accuracy.

Algorithmic Approaches (primarily within Vision-Based): * Two-Stage Methods: Involves a "Hypothesis Generation" (HG) step to propose candidate regions (ROIs) followed by a "Hypothesis Verification" (HV) step to confirm vehicle presence. HG methods include appearance-based (color, edge, corner, symmetry, texture, shadow, lights), motion-based (frame difference, background modeling, optical flow), and stereo-based techniques. HV methods include template-based and classifier-based approaches. * Deep-Learning Based Methods: Covers both two-stage (e.g., Faster R-CNN) and one-stage (e.g., YOLO, SSD) neural network detectors, highlighting their advantages in speed and accuracy.

The paper also reviews simulation platforms and datasets crucial for algorithm development and validation.

Datasets

The paper summarizes several datasets relevant to UGVs, including:

  • Apollo: Expressway data under various weather conditions, including images, labels, and sensor data.
  • BDDV: Road/weather/lighting conditions data with images, labels, and sensor data.
  • Ford: Downtown, loop closure, and campus data with lidar scans, images, and sensor logs.
  • JAAD: Urban data with monocular camera, textual, and bounding box annotations.
  • Karlsruhe labeled objects: Urban daylight data with images and object labels/orientation.
  • Karlsruhe stereo: Urban/rural daylight data with stereo camera, GPS/IMU data, and calibration.
  • KITTI: Urban/rural/highway data with multiple cameras, Lidar, GPS/IMU, and 3D bounding box annotations.
  • Malaga: Urban/highway loop closure data with stereo cameras, Lidar, and GPS/IMU.
  • MVD: Various road/weather/lighting conditions data from diverse devices.
  • Stanford: Urban campus intersection data with Lidar and GPS/IMU.
  • Udacity: Sunny/overcast daylight data with monocular camera, Lidar, and GPS/IMU.
  • CityScapes: Urban daytime data with stereo camera and GPS/IMU.
  • H3d-HRI-US: Urban traffic condition data with multiple cameras, Lidar, and GPS/IMU.
  • nuScens: Urban/various weather/traffic condition data with Lidar, cameras, radars, and GPS/IMU.

Key Insights and Ideas

  • Sensor Diversity is Key: No single sensor is perfect for all conditions. Vision sensors excel at capturing color and texture but struggle with depth and adverse weather. Lidar provides accurate depth but lacks color and texture. Radar offers robustness to weather but has lower resolution. Infrared is useful for low-light but has low resolution. Fusion is often necessary for comprehensive perception.
  • Deep Learning Dominance: Deep learning, particularly convolutional neural networks (CNNs), has become the de facto standard for many vehicle detection tasks, offering significant improvements in accuracy and efficiency over traditional methods. Both two-stage and one-stage detectors are widely used.
  • The Importance of Simulation and Datasets: The development and validation of UGV algorithms heavily rely on realistic simulation platforms (e.g., CARLA, AirSim, Gazebo) and comprehensive datasets (e.g., KITTI, nuScenes) that capture diverse scenarios and sensor modalities.
  • Challenges Remain: Despite advancements, challenges persist in areas like detecting occluded vehicles, handling extreme weather and illumination conditions, achieving real-time performance at high speeds, and detecting special vehicle types.

Key Results

The paper provides a systematic review and comparison of numerous vehicle detection methods across various sensors and algorithmic approaches. While specific quantitative results are presented within the context of individual cited works (e.g., accuracies ranging from 70% to over 99% for different methods and datasets), the primary "result" of this paper is its comprehensive synthesis and organization of this vast body of research. It highlights the trade-offs between different sensor modalities and algorithmic techniques in terms of accuracy, computational cost, robustness, and real-time performance.

Limitations and Future Work

Limitations: * The paper acknowledges that while sensor fusion can improve performance, it increases complexity, computing time, and sensor costs. * For simple driving scenes, high-precision detection might be achievable with a single sensor, making fusion potentially resource-intensive. * The paper notes that many current methods are limited to specific scenarios or vehicle types.

Future Work: The authors identify several promising research directions:

  • Sensor-Based Anomaly Detection: Developing robust methods to detect sensor failures or anomalies for improved safety.
  • Multi-Mode Sensor Fusion: Creating adaptive sensor fusion strategies that switch modes based on driving conditions to maximize resource utilization.
  • Special Vehicle Inspection: Focusing on detecting specialized vehicles (ambulances, fire engines, military vehicles) by constructing feature databases and datasets.
  • Vehicle Detection under High Speed: Improving the real-time performance of algorithms to ensure accurate perception during high-speed driving.

Summary updated

Abstract

Unmanned ground vehicles (UGVs) have great potential in the application of both civilian and military fields, and have become the focus of research in many countries. Environmental perception technology is the foundation of UGVs, which is of great significance to achieve a safer and more efficient performance. This article firstly introduces commonly used sensors for vehicle detection, lists their application scenarios and compares the strengths and weakness of different sensors. Secondly, related works about one of the most important aspects of environmental perception technology-vehicle detection-are reviewed and compared in detail in terms of different sensors. Thirdly, several simulation platforms related to UGVs are presented for facilitating simulation testing of vehicle detection algorithms. In addition, some datasets about UGVs are summarized to achieve the verification of vehicle detection algorithms in practical application. Finally, promising research topics in the future study of vehicle detection technology for UGVs are discussed in detail.
BibTeX
@inproceedings{liu2021review,
  author = {Qi Liu and Zirui Li and Shihua Yuan and Yuzheng Zhu and Xueyuan Li},
  title = {{Review on Vehicle Detection Technology for Unmanned Ground Vehicles}},
  booktitle = {Sensors},
  year = {2021},
  doi = {10.3390/s21041354},
}
Figure