Deep Learning for Efficient Adverse Weather Image Processing in Autonomous Driving

Engineering notes

Applied to popular CNN backbones, SSP-X achieves state-of-the-art performance in weather classification and object detection, with 60-90% reductions in computational complexity. Additionally, a multi-label weather classification framework is proposed that can simultaneously capture multiple co-occurring weather conditions present in images to improve robustness and realism in ADS perception.All contributions are validated through extensive experiments on benchmark and domain-specific datasets and further demonstrated in a novel perception module integrating weather classification, noise removal, and object detection within an AD perception system.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Autonomous Driving Systems (ADSs) are becoming increasingly important to daily life due to their ability to see, interpret, and act in complex environments. However, their reliability and safety critically depend on robust performance under a wide range of real-world scenarios. One of the biggest challenges comes from adverse weather conditions, which can seriously affect visual perception tasks. This thesis addresses that challenge by improving the robustness of Convolutional Neural Networks (CNNs), both architecturally and functionally, to make them more reliable for visual detection and classification in ADSs operating under such complex conditions.The core technical contribution is the EXtended Sparse Split Parallelism (SSP-X) design framework, a novel CNN design paradigm. SSP-X builds on the original SSP framework by adding depthwise separable convolutions for model size reduction, optimised incorporation of normalisation layers, deactivating redundant instances to enhance efficiency, and the first systematic exploration of the Dynamic Tanh layer in CNNs, showing that accuracy can be maintained while improving efficiency. Applied to popular CNN backbones, SSP-X achieves state-of-the-art performance in weather classification and object detection, with 60-90% reductions in computational complexity. Moreover, this thesis introduces a new evaluation methodology for CNN-based approaches designed for ADSs that is adaptable to any imaging task. It enables a realistic assessment of model readiness for deployment in real-world ADSs, setting a new standard for how weather classification models should be tested and compared. Additionally, a multi-label weather classification framework is proposed that can simultaneously capture multiple co-occurring weather conditions present in images to improve robustness and realism in ADS perception.All contributions are validated through extensive experiments on benchmark and domain-specific datasets and further demonstrated in a novel perception module integrating weather classification, noise removal, and object detection within an AD perception system. Tests on a small-scale embedded platform confirm real-time feasibility of the proposed contributions.

Links and sources

Comments