Optimizing Semantic Segmentation for Autonomous Vehicle Scene Understanding in Unstructured Indian Traffic through Reinforced Active Learning

Abstract

Autonomous vehicles (AVs) offer a radical leap in transportation, delivering safer and more efficient mobility options. The capacity to interpret complicated surrounding traffic scenarios in real-time is central to their effectiveness. Scene awareness, especially semantic segmentation, is vital in allowing AVs to successfully comprehend and navigate their environments. However, limited labelled data availability and dataset biases restrict the effectiveness of semantic segmentation models, especially in specific contexts such as Indian driving scenarios. This study presents a novel approach employing reinforced active learning to overcome the aforementioned difficulties. Reinforced active learning integrates reinforcement learning into the active learning framework, allowing the model to select samples for annotation based on model operations and uncertainty estimation. By augmenting the segmentation model with annotation effort, our approach enhances performance in real-world driving scenarios in India. Rigorous testing and validation on the Indian Driving Dataset (IDD) demonstrate improvements in segmentation precision and effectiveness compared to training methods. Reinforcement Active Learning (RAL) using Inception-Unet outperforms Inception-Unet models trained solely on labeled data (DL), achieving a score of 0.615. However, it falls slightly behind the performance of Inception-Unet models trained on fully labeled datasets (DF). Our findings indicate that reinforced learning excels over strategies in selecting samples and substantially boosts segmentation accuracy.