SPIF: A Spatio-Temporal Polarity Interaction Filter for Reliable Event Selection

Abstract

Event cameras provide high temporal resolution and sparse asynchronous output for small-target monitoring. However, distant weak targets often generate sparse and fragmented events that are easily obscured by responses from background structures and sensor background activity (BA) noise. This paper proposes a Spatio-Temporal Polarity Interaction Filter (SPIF), which assigns a reliability score to each incoming event. SPIF constructs weighted neighborhood support from temporal proximity, spatial distance, and polarity relationships, and uses the recent firing history of the center pixel to discount unreliable evidence caused by persistent activation. On the complete test split of the event-based unmanned aerial vehicle (EV-UAV) benchmark, SPIF achieves a macro-averaged F1 score (macro-F1) of 0.7732 and a target retention rate of 0.8896, exceeding the corresponding values of 0.7583 and 0.6740 obtained by the supervised EV-SpSegNet. Driving and the ED24 real-world denoising dataset further evaluate separation between valid events and BA noise. SPIF obtains the highest area under the receiver operating characteristic curve (AUC) on Driving at 3–10 Hz/pixel and under all evaluated ED24 settings. The C++ implementation processes 6.72 million events per second on the tested CPU.