Event Camera Motion Detection via Grid Partitioning and Persistent Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Event cameras struggle to efficiently filter, cluster, and track objects in motion when faced with highly dynamic backgrounds that produce significantly more intensity changes than the objects of interest, leading to excessive computation and reduced accuracy in object detection and tracking.
Innovation Solution
The method involves partitioning the event camera's field of view into grids, filtering events within each partition to reduce noise, buffering filtered events, clustering events based on spatio-temporal proximity, sorting and tracking clusters to identify objects of interest, and outputting detection results in real-time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If event cameras capture all intensity changes in the field of view, then motion detection capability is improved, but computational overhead increases significantly due to highly dynamic backgrounds producing orders of magnitude more events than objects of interest
Solution Approach 1:
The patent divides the field of view into multiple partitions or regions of interest. By segmenting the scene, the system processes events only in relevant areas rather than the entire field of view, significantly reducing computational overhead while maintaining motion detection capability for objects of interest.
Solution Approach 2:
The patent applies different processing strategies to different regions of the field of view. Regions containing objects of interest receive full event processing, while regions with only dynamic background receive reduced processing or are excluded entirely. This local differentiation reduces overall computational load while preserving detection accuracy where needed.
2Quantity of substance
If random sampling of events is applied across the entire field of view, then data redundancy is reduced, but detection accuracy decreases in areas not of interest while maintaining unnecessary processing
Solution Approach 1:
The patent segments the field of view into partitions and applies selective processing to each. By identifying which partitions contain objects of interest, the system reduces data redundancy in background areas while maintaining full detection accuracy in regions containing targets, avoiding the uniform reduction applied by random sampling.
Solution Approach 2:
The patent applies different sampling or processing densities to different spatial regions. Regions with objects of interest maintain high event processing fidelity, while background regions use reduced processing. This local quality approach preserves detection accuracy where needed while reducing redundancy where unnecessary, unlike uniform random sampling.
3Device complexity
If grouping is based on flow vector, then motion analysis is simplified, but object isolation fails when objects of interest are grouped with dynamic background due to similar motion patterns
Solution Approach 1:
The patent first segments the field of view into partitions before applying flow-based grouping. By restricting clustering operations to specific partitions containing objects of interest, the system maintains the computational simplicity of flow vector grouping while preventing spurious grouping with dynamic background that occurs when processing the entire field of view.
Solution Approach 2:
The patent extracts or isolates events from specific partitions containing objects of interest before applying flow vector grouping. This extraction step removes dynamic background events from the grouping process, allowing flow-based motion analysis to remain computationally simple while improving object isolation accuracy by preventing mixing with background motion.
4Productivity
If deep learning approaches with GPU acceleration are used, then real-time processing is achieved, but power consumption increases significantly
Solution Approach 1:
The patent segments the processing workload by partition and prioritizes processing of partitions containing objects of interest. This segmentation allows the system to achieve real-time processing for critical regions using more energy-efficient methods, while reducing or skipping processing in less critical background regions, thereby reducing overall power consumption compared to uniform deep learning processing.
Solution Approach 2:
The patent applies different processing quality levels to different regions. Regions with objects of interest receive high-quality real-time processing, while background regions receive reduced processing or are excluded. This local quality approach achieves real-time performance where needed without the excessive power consumption of applying deep learning uniformly across the entire field of view.
Data Source
AI summary
The invention is a system and method for efficient filtering, clustering, tracking and persistent motion detection for event cameras. More particularly, embodiments of the present invention combine compatible asynchronous event filtering and clustering techniques in a computationally efficient way to identify and track objects in the field of view that exhibit persistent motion characteristic of marine vessels, cars, aircraft, and other vehicles while ignoring dynamic background elements that move in a more erratic fashion.