Abstract
This paper presents a vision-based obstacle detection system utilizing a novel Fisheye Lens Inverse Perspective Mapping (FLIPM) method. By deriving new mapping equations, the system transforms distorted fisheye images into undistorted, top-down remapped views. Obstacle detection is achieved by identifying vertical edges, which project as radial lines in the remapped image. The system integrates profile searching and temporal IPM difference imaging to extract features, which are then processed via polar histograms to determine obstacle positions while filtering noise. Experimental results demonstrate the system's effectiveness in challenging conditions such as rain and night, providing a wide 125° field of view for vehicle safety applications.
Introduction
Standard vehicle-mounted cameras typically possess a narrow field of view (FOV), often between 30° and 40°, leaving significant blind spots around the vehicle. While fisheye lenses provide a wide-angle solution, they introduce severe radial distortion—the "barrel" effect—that complicates traditional computer vision tasks.
Traditional Inverse Perspective Mapping (IPM) techniques, which assume a pinhole camera model to create a "bird’s-eye view," fail to correct these distortions. Furthermore, standard IPM often maps straight horizontal lines into arcs, leading to inaccuracies. This research introduces the FLIPM model to provide a unified mathematical framework that handles fisheye distortion and ensures robust obstacle detection across a massive 125° FOV.
System Architecture and the FLIPM Model
The proposed system follows a multi-stage pipeline to transform raw video input into actionable obstacle information.
1. The FLIPM Equations
The core of the system is the FLIPM model, which combines a spherical projection model with perspective transformation. Unlike previous iterations, these equations ensure that horizontal lines in the physical world remain horizontal in the remapped image, facilitating more accurate spatial reasoning.
2. Feature Extraction and Dual-Path Processing
The system identifies obstacles by searching for vertical edges. In a bird's-eye view, any object with physical height (such as a vehicle or pedestrian) projects as a line pointing directly toward the camera's ground-projection point. To extract these features, the system uses two parallel paths:
- Profile Searching: This path processes static scenes using sharpening, Sobel edge detection, and modified thinning algorithms to find object boundaries.
- Temporal Difference: To detect moving obstacles, the system compares the current frame with a spatially shifted version of the previous frame. This "pseudo-stereo" approach highlights non-planar objects while effectively canceling out flat road markings.
Obstacle Localization via Polar Histograms
A key insight of this work is the Radial Line Property: in a correctly remapped IPM image, the vertical edges of a 3D object will always lie on a radial line passing through the camera's projection point.
By plotting extracted edges on a polar histogram, the complex 3D detection problem is reduced to a 1D peak-searching task. Obstacles appear as dominant peaks in the histogram. To improve accuracy, the authors implemented "Trapezoid Filtering" to eliminate false positives caused by lane markings, which exhibit a distinct trapezoidal signature in the polar domain.
Experimental Results
The system was evaluated using custom-captured video sequences across various environments, including parking lots, urban streets, and campus settings.
- Wide Field of View: The FLIPM system achieved a 125° FOV, significantly wider than the 28°–48° FOV of contemporary systems like GOLD.
- Real-Time Performance: Processing averaged 64.46 ms per frame on a 1.6 GHz CPU, enabling a near real-time frame rate of approximately 15 FPS.
- Environmental Robustness: The system demonstrated high reliability in challenging conditions, such as nighttime and heavy rain. By utilizing temporal differences, it successfully ignored light reflections on wet pavement that often trigger false detections in simpler IPM models.
Conclusion
The FLIPM model provides a robust solution for wide-angle obstacle detection by mathematically correcting fisheye distortion and leveraging the geometric properties of vertical edges. While highly effective for objects with clear vertical profiles, future work may involve integrating additional sensors (e.g., radar or ultrasonic) to detect low-profile or slanted obstacles and improving performance in extremely low-contrast environments.
Summary updated