Robust Traffic Vehicle Detection under Real-World Conditions with Misclassified Vehicles Minimization and Weighted Box Fusion

Traffic vehicle detection is a vital component of intelligent transportation systems, which encompass technologies used in modern automotive applications such as autonomous driving, traffic monitoring, and vehicle safety systems. However, real-world conditions—especially the frequent misclassification of pedestrians as motorbikes in crowded urban environments—pose significant challenges to achieving high detection accuracy. To tackle these challenges, we employed two primary techniques: a misclassified vehicle minimization strategy to reduce false positive and false negative predictions, particularly addressing the common issue of pedestrians being misclassified as motorbikes, and Weighted Box Fusion to refine object detection results by aggregating predictions from multiple models. Our proposed pipeline integrates state-of-the-art object detection architectures with targeted post-processing strategies to correct classification errors and enhance prediction robustness through model fusion. Experimental evaluations show that our method significantly improves detection performance in complex traffic scenarios, securing the 2<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">nd</sup> place in the SoICT Hackathon 2024 – "Traffic Vehicle Detection" Track, where participants were required to detect various vehicle types, including motorbikes, cars, buses, and trucks.