FNO-LSTM: Context-Aware Feature Fusion for Edge Camera Event Summarization with Reliability Gating

Abstract

Edge cameras can find many local incidents, but sending full streams or disconnected alarm clips over constrained links overloads these links and gives operators redundant evidence. This paper proposes a context-aware event summarization model that combines Fourier Neural Operators and Long Short-Term Memory networks. The operator learns the cross-window temporal dynamics in a compressed spectral space, and the recurrent branch maintains the order and duration of events. Scene state, object interaction, motion intensity, illumination and device load are fused by a reliability-gated context module prior to diversity-aware key-clip selection. A four-camera edge testbed was evaluated using 18,420 annotated events from traffic, campus, warehouse and perimeter scenes. At a 12-clip budget, the proposed model reached 0.842 F1, 0.781 mean average precision and 0.736 temporal intersection over union, and outperformed the best temporal convolution baseline by 4.7, 4.3 and 5.1 percentage points, respectively. Reduced the transmitted video volume by 91.6%, maintained 27.8 frames per second at a 15 W edge module, and kept the 95th-percentile summary delay at 428 ms. Ablation experiments show that spectral modelling and reliability-gated context fusion are complementary; they perform well under light-variation and partial-occlusion conditions. Based on the above results, operator-recurrent fusion can generate compact, ordered and operationally feasible camera summaries without continuous cloud inference.