Wavelet-enhanced transformers for real-time event-based object detection in low-light environments

Neuromorphic event cameras exhibit superior performance in low-light environments owing to their high dynamic range and microsecond-level temporal resolution. However, the intensive background activity (BA) noise inherent in low-illumination event streams significantly compromises detection accuracy. Conventional strategies predominantly utilize a decoupled pipeline where denoising and detection are treated as independent stages, often resulting in irreversible information loss and suboptimal performance due to mismatched optimization objectives. To address these challenges, we propose an end-to-end Wavelet-Enhanced Transformer detection network (WE-DETR). The proposed framework integrates the wavelet transform directly into the feature extraction hierarchy to adaptively suppress stochastic noise and enhance multi-scale structural representations from event accumulations. This is followed by a Transformer-based detection head designed to model global spatiotemporal dependencies for precise object localization and recognition. Extensive evaluations on three benchmark datasets–DSEC, Gen1, and 1Mpx–demonstrate that WE-DETR consistently outperforms state-of-the-art methods. Furthermore, hardware deployment on the NVIDIA Jetson Orin NX achieves real-time inference at approximately 35 FPS using TensorRT FP16, validating the framework's suitability for resource-constrained embedded platforms. Field experiments in diverse outdoor scenarios under illumination as low as 2.7 lux further confirm the robustness and practical utility of WE-DETR in extremely low-light environments.