Non-Linear DVS Event Mapping for Faint Feature Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing Dynamic Vision Sensor (DVS) technologies face challenges in accurately capturing fast scene changes and recovering lost information between successive frames, which affects the accuracy of computer vision applications like Simultaneous Localization and Mapping (SLAM).
Innovation Solution
The proposed solution involves a method of imaging DVS events that includes non-linearly incrementing motion-compensated pixel locations, representing events as a weighted linear combination of polar and non-polar components, and performing event-density based adaptive normalization to enhance contrast and feature detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If DVS events are accumulated using linear incrementing of pixel intensity, then the processing is simple and fast, but faint features are lost and contrast is insufficient
Solution Approach 1:
The patent applies parameter changes by transitioning from linear to non-linear incrementing of pixel intensity values. Specifically, it uses exponential incrementing where each subsequent event adds a progressively larger value (e.g., 1, 2, 4, 8, 16...) rather than a constant increment. This parameter change enhances the contrast between regions with different event densities, making faint features more detectable while maintaining computational efficiency through bitwise operations.
Solution Approach 2:
The patent employs curvature in the form of non-linear (exponential) scaling of pixel intensities rather than linear scaling. This curvature in the intensity distribution amplifies the representation of regions with higher event densities more aggressively, creating a contrast-enhanced representation that preserves faint features while maintaining processing simplicity.
2Measurement precision
If contrast enhancement is applied to boost faint features, then feature detection improves, but noise is also amplified
Solution Approach 1:
The patent applies local quality by performing contrast enhancement locally at each pixel position based on its specific event density. The non-linear incrementing scheme automatically adapts to local conditions: pixels with high event density receive larger cumulative values, while pixels with low density receive smaller values. This localized approach enhances features where needed without uniformly amplifying noise across the entire image.
Solution Approach 2:
The patent uses dynamic contrast enhancement where the increment value changes dynamically based on the event density at each location. The exponential scaling factor adapts to the local event rate, providing stronger enhancement in high-density regions (where features are more prominent) and milder enhancement in low-density regions (where noise would be more problematic).
3Measurement precision
If motion compensation is applied to improve accuracy, then feature tracking precision improves, but computational cost increases
Solution Approach 1:
The patent replaces complex mechanical-style motion compensation calculations with a simpler event-driven accumulation approach. Instead of tracking individual feature points through complex geometric transformations, the system accumulates events at pixel locations with non-linear weighting, which implicitly performs motion compensation through the temporal accumulation process. This substitution reduces computational overhead while maintaining precision.
4Speed
If DVS is used to capture fast scene changes, then dynamic range and data rate improve, but feature tracking accuracy deteriorates due to lost information
Solution Approach 1:
The patent applies preliminary action by accumulating events over time with non-linear weighting before feature detection and tracking. This pre-accumulation process with exponential incrementing prepares the data by enhancing contrast and preserving faint features that would otherwise be lost in high-speed capture. The preliminary enhancement ensures that subsequent tracking operations work with improved quality data, compensating for the information loss inherent in high-speed DVS capture.
Data Source
AI summary
The disclosure relates to methods and systems for non-linear mapping of motion-compensated DVS events to DVS images. In the non-linear mapping of motion-compensated DVS events to DVS images, the current pixel increments depend on the existing number of accumulated events at that location. Further, the initial events are given a larger weightage to preserve the tracked features which are used for bundle adjustment. Further, disclosure relates to methods and systems for representing polarity in single channel DVS Frame. As the polarity adds additional constraints on feature matching and therefore accurate optical flow is seen in the image. Moreover, disclosure relates to methods and systems for using event-density as a measure of input contrast after DVS event accumulation and this is used to determine the target range for contrast stretching.