自适应时空注意力机制改进神经辐射场方法研究

打开文本图片集

中图分类号：TP391.4;G434 文献标识码：A 文章编号：1006-8228（2025）11-19-07

Abstract：High-precisionreconstructiondynamicscenesisgreatsignificanceinfieldssuchasautonomousdrivingand augmentedreality.ThispaperproposesadynamicNeuralRadianceFieldreconstructionandmotioncompensationmethodbasedon eventcameras，caledSTAA-NeRF.Thismethoddesignsadiferentiableevententropy-guidedadaptivespatio-temporalaention module（DEE-SA），whichcanadaptivelyaggegatedynamicfeaturesinmultipleframeseventstreams，improvingthemodeling abilityforhigh-speedmovingareas.Atthesametime，itintroducesanevent-raydiferentialcouplingmodel（ERDCM），which realizessamplingsetanddensityadjustmentthrougheventgradients，therebytheoreticallyensuringtheconsistencymotion compensationandgeometricmodeling.Experimentalresultsshowthatinrealandsynthetichigh-speedscenes，STAA-NeRF outperformsDeblur-NeRFwitha2.6dBincreaseinPSNR，a0.1increaseinSIM，andaO.04reductioninLPIPS，and’til maintainsrobustperformanceinlow-lightandhigh-spedmotionenvironments.Theresearchresultsprovideanewtechnicalpath forreal-time perception and high-quality reconstruction dynamic vision systems.

Keywords：Dynamic Scene;Neural Radiance Field； Event Camera;Adaptive Spatiotemporal Attention

0引言

动态场景的高质量时空重建在计算机视觉与智能感知领域具有重要的研究价值。（剩余9996字）