Background Subtraction for Dynamic Lighting in Autonomous Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing background subtraction techniques are ineffective in handling variant illuminating conditions, leading to false recognition of foreground objects in autonomous driving scenarios, particularly in dynamic environments with changing light conditions.
Innovation Solution
A method that collects baseline images under different illuminating conditions to derive a normality model, projects a normality background image, and applies an illumination-invariant color space transformation to improve consistency in object recognition, using a Hadamard product to refine foreground pixel identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional background subtraction techniques are used, then the system is simple to implement, but false recognition of foreground objects occurs under variant illuminating conditions
Solution Approach 1:
The system performs preliminary actions by collecting baseline images under different illuminating conditions before actual foreground detection. A normality model is derived in advance from these baseline images, capturing the statistical characteristics of background pixels under various lighting conditions. This pre-computation enables the system to adapt to illumination changes without increasing real-time processing complexity.
Solution Approach 2:
The system changes parameters by transforming images into an illumination-invariant color space and dynamically adjusting the normality model based on detected illuminating conditions. The standard deviation threshold is adaptively modified according to the current illumination state, allowing the background subtraction to remain accurate across varying lighting conditions without requiring a completely different system architecture.
2Measurement precision
If a global threshold value is used for inter-frame difference, then the method is computationally simple, but it fails to detect objects when background assumptions are violated
Solution Approach 1:
Instead of using a single global threshold, the system applies local quality by computing a normality model with per-pixel mean and standard deviation values. Each pixel has its own adaptive threshold based on the local background characteristics captured in the normality model. This allows the detection sensitivity to vary locally across the image, improving detection precision while maintaining computational efficiency through vectorized operations.
3Adaptability or versatility
If baseline images are collected under different illuminating conditions, then the system adapts to lighting changes, but the processing time and computational load increase
Solution Approach 1:
The system performs preliminary action by collecting and processing baseline images in advance to build the normality model before actual operation. This pre-computation phase captures the statistical properties of the background under various illuminating conditions, so that during real-time operation, the system only needs to compare current frames against the pre-built model, significantly reducing processing time.
Solution Approach 2:
The system creates a copy of the background information in the form of a normality model that encapsulates the statistical characteristics of baseline images. Instead of re-processing baseline images during operation, the system uses this compact copied representation (mean and standard deviation maps) for rapid comparison, maintaining adaptability while minimizing computational load during time-critical operations.
Data Source
AI summary
A method and an apparatus for background subtraction highly applicable in autonomous driving scenarios are described. The method involves a reduction of illumination effects by constructing a normality background image from a normality model based on a plurality of baseline images taken under different illuminating conditions. A subtracted image is obtained by subtracting the normality background image from a scene image pixel-wise (i.e., pixel-by-pixel). The scene image may contain one or more foreground objects. The foreground objects are identified by highlighting the pixels in the subtracted image whose intensity is more than a predetermined standard deviation in the normality model. An illumination-invariant color space transformation algorithm may optionally be utilized to further reduce the variant illumination effects.