Core Concepts
The proposed EDA approach effectively and explicitly handles the event data association and fusion problem by performing robust multi-structural model fitting on event data to estimate accurate event trajectories.
Abstract
The paper proposes a novel Event Data Association (EDA) approach to address the fundamental event data association problem for event-based object tracking.
The key steps are:
Asynchronous event fusion: The sequential retinal events are asynchronously fused into different sets based on the information entropy of the accumulated events, to leverage the asynchronous nature of event data.
Deterministic model hypothesis generation: A deterministic strategy is introduced to effectively generate model hypotheses in the spatio-temporal domain, which represent event trajectory candidates.
Robust model hypothesis selection: A two-stage weighting algorithm is proposed to robustly weigh and select the true event trajectory models from the generated hypotheses, through multi-structural geometric model fitting. An adaptive model selection strategy is also presented to automatically determine the number of true models.
Event data association and fusion: The selected true event trajectory models are used to associate and fuse the event data, without being affected by sensor noise and irrelevant structures.
The proposed EDA is extensively evaluated on object tracking tasks, demonstrating superior performance over state-of-the-art event-based and conventional tracking methods, especially under challenging conditions like high speed, motion blur, and high dynamic range.
Stats
The paper does not provide any specific numerical data or statistics. The focus is on the algorithmic contributions for event data association and fusion.
Quotes
"Event-based approaches, which are based on bio-inspired asynchronous event cameras, have achieved promising performance on various computer vision tasks."
"Since data association is the key to various fundamental computer vision tasks (such as object tracking), the problem of event-based data association needs further investigation."
"We explicitly formulate the event-based data association problem as a 3D event trajectory estimation problem in the spatio-temporal domain."