Multi-task driven robust object detection for low-light traffic scenes

Abstract

The rapid advancement of intelligent driving and transportation systems has made object detection in traffic scenarios a key research area. However, nighttime traffic images often suffer from poor illumination, low visual quality, and high noise, which degrade the performance of conventional detection methods. To overcome these challenges, this paper proposes a deep neural network framework based on multi-task learning, integrating object detection with low-light image enhancement to improve detection accuracy in complex nocturnal environments. Unlike traditional methods where image enhancement is a separate preprocessing step, our end-to-end framework employs a shared feature extraction strategy. A unified backbone network extracts features that are used simultaneously by both an image enhancement sub-network and an object detection sub-network. By jointly optimizing the detection and enhancement loss functions, this approach enables implicit feature enhancement, improving task performance. To address the difficulty of acquiring paired normal-light reference images in real traffic conditions, we use a reference-free image enhancement network. Additionally, a dynamic weight adjustment mechanism is introduced to balance the loss function weights during training. Experimental results on both synthetic and real-world datasets demonstrate that the proposed method significantly outperforms existing algorithms, highlighting the effectiveness and advancement of this multi-task learning framework.