Uncertainty-Aware Deep Learning Models for Robust Realtime Object Recognition in Dynamic Computational Vision

Abstract

The challenge of real-time object detection in dynamic environments is complicated by issues like noise, occlusion, changes in illumination, and ambiguity of object boundaries. Non-Bayesian models of deep learning tend to be confident about their predictions. Such approach is dangerous since it renders these models inappropriate for use in safety-critical applications such as self-driving vehicles, surveillance and robotics. In this paper, an uncertainty-aware deep learning method is suggested which can be applied to real-time object detection in dynamic computational vision environments. This method combines a lightweight detector based on YOLO architecture, Monte Carlo dropout and uncertainty estimation via entropy measure to account for both aleatoric and epistemic uncertainties. The approach is expected to increase robustness to challenges including occlusion, motion blur and illumination variation. Experimental results obtained on COCO and KITTI data sets show that the proposed model reaches mAP of 88.9%, that is, 6.8% better compared to baseline YOLO and CNNs models. False positives are reduced by 12.3% and ECE score is increased by 9.5%. The model runs at 38 FPS which ensures its real-time operation. The results confirm that uncertainty-aware reasoning significantly enhances prediction reliability and interpretability in object detection systems.