Abstract
Event cameras have shown great potential for autonomous driving due to their high temporal resolution and wide dynamic range. However, most existing event-based object detectors rely on LSTM-based temporal aggregation, which remains insufficient in highly dynamic scenes. In particular, conventional LSTMs depend on a single global memory to model temporal dependencies, making them vulnerable to abrupt variations in target-related event patterns and thus limiting robustness. Moreover, their implicit temporal modeling is not well suited to capturing motion dynamics explicitly, which often leads to inaccurate object localization. To address these issues, we propose EDeMM, a novel event-based object detection framework with enhanced memory and motion modeling capabilities. Specifically, we design a Double-Phase LSTM (DPLSTM) consisting of Outer and Inner recurrent units to jointly capture global temporal dependencies and local temporal variations, thereby improving memory robustness under dynamically changing inputs. We further introduce a Motion-Enhancement Module (MEM) to explicitly strengthen motion trend representation and improve localization accuracy in highly dynamic scenarios. Extensive experiments on the GEN1 and 1Mpx datasets demonstrate that EDeMM achieves superior detection performance and robustness under challenging real-world conditions.
Funding Information
  • National Natural Science Foundation of China (62501435)
  • National Natural Science Foundation of China (62402366)

This publication has 13 references indexed in Scilit: