Motion blur in traffic perception causes direction-dependent spectral degradation that severely attenuates boundary, contour, and texture details needed for reliable vehicle detection. Existing methods that rely on restoration-first pipelines often introduce reconstruction artifacts and add computational latency to visual processing. Meanwhile, detector-side robustness approaches frequently depend on fixed degradation distributions during training without explicitly correcting blur-induced spectral discrepancies.
The proposed framework combines a progressive motion-blur learning strategy with multi-scale spectral feature refinement. This architecture addresses frequency loss by directly integrating spectral correction into the feature extraction process of an object detection network.
During the training phase, PBSA employs a non-learnable curriculum based on point-spread functions to simulate realistic motion degradation. This curriculum progressively expands the admissible motion-kernel range while simultaneously increasing the probability of blur exposure. As a result, the training procedure generates paired sharp and blurred image observations of increasing degradation difficulty, which are then processed through a shared YOLOv8n detector.
To restore attenuated frequency information, the architecture applies scale-specific and channel-shared real-valued spectral gates at the P3, P4, and P5 feature levels. These spectral gates operate directly on two-dimensional real Fourier coefficients extracted from the feature representations.
Beyond gating, the framework aligns the amplitude spectra of the compensated blurred features with stop-gradient sharp-reference spectra. Both the blurred and sharp reference branches maintain active detection supervision throughout training to ensure reliable localization performance.
Evaluation on benchmark datasets demonstrated marked improvements under synthetic motion blur conditions compared to the baseline YOLOv8n detector. Under heavy motion blur, PBSA achieved an mAP50 of 0.818 and an mAP50–95 of 0.583 on the UA-DETRAC dataset, compared to 0.714 and 0.468 for the baseline. On the UAVDT benchmark, PBSA recorded an mAP50 of 0.286 and an mAP50–95 of 0.163, surpassing the baseline scores of 0.171 and 0.094. The researchers noted that performance gains observed on clear images may partly reflect the regularization effect of training.
The framework was further tested on real-world degradation using the 500-image real-motion-blur subset of the BDD100K driving dataset. On this benchmark, PBSA achieved an mAP50 of 0.556, substantially improving upon the 0.446 mAP50 recorded by the baseline detector. The resulting detection model contains 4.4 million parameters and operates at a processing speed of 109.7 frames per second, confirming its practical efficiency for real-time traffic perception applications.
Subject of Research: Robust vehicle detection under motion blur using progressive spectral alignment
Article Title: PBSA: progressive blur-aware spectral alignment with multi-scale frequency refinement for robust vehicle detection
Article References: Li, X., Li, R., & Sun, X. (2026). PBSA: progressive blur-aware spectral alignment with multi-scale frequency refinement for robust vehicle detection. Journal on Advances in Signal Processing. https://doi.org/10.1186/s13634-026-01382-4
Image Credits: AI Generated
DOI: 10.1186/s13634-026-01382-4
Keywords: vehicle detection, motion blur, spectral alignment, Fourier coefficients, progressive blur curriculum, spectral gating, point spread function, YOLOv8n, UA-DETRAC, UAVDT, BDD100K, real-time traffic perception, frequency refinement, amplitude spectra, direction-dependent degradation, feature refinement
Cite Scienmag News
Scienmag. (November 19, 2026). Spectral Alignment Framework Improves Vehicle Detection Accuracy Under Severe Motion Blur. https://scienmag.com/spectral-alignment-framework-improves-vehicle-detection-accuracy-under-severe-motion-blur/
Scienmag. "Spectral Alignment Framework Improves Vehicle Detection Accuracy Under Severe Motion Blur." Scienmag, 19 November 2026, https://scienmag.com/spectral-alignment-framework-improves-vehicle-detection-accuracy-under-severe-motion-blur/. Accessed 19 November 2026.
Scienmag. "Spectral Alignment Framework Improves Vehicle Detection Accuracy Under Severe Motion Blur." Scienmag. November 19, 2026. https://scienmag.com/spectral-alignment-framework-improves-vehicle-detection-accuracy-under-severe-motion-blur/








