TCI-Net: Complexity-Driven Domain Adaptation for Robust ADAS Perception in Unstructured Traffic

Perception models for autonomous driving are largely trained on Western datasets and degrade significantly when deployed in emerging markets with heterogeneous and unstructured traffic. Environments such as India and Southeast Asia exhibit mixed vehicle types, informal driving behavior, weak lane discipline, and frequent occlusions, leading to a 30–40% drop in detection accuracy for models trained on benchmarks like NuScenes and KITTI.

This paper presents two deployment‑focused contributions. First, we introduce a Traffic Complexity Index (TCI), a real‑time metric that quantifies scene chaos based on object density, motion unpredictability, road structure variance, and occlusion patterns. Second, we propose a lightweight domain adaptation framework using feature alignment and self‑training to adapt existing perception models without full retraining. The approach is designed for automotive‑grade hardware, achieving sub‑50 ms latency on platforms such as Qualcomm SA8295 and NVIDIA Orin.

Experiments on the India Driving Dataset show a 34% improvement in mean average precision with under 8% computational overhead, demonstrating a scalable path to robust ADAS perception in emerging markets.