ANNA: A Deep Learning Based Dataset in Heterogeneous Traffic for Autonomous Vehicles
Mahedi Kamal, Tasnim Fariha, Afrina Kabir Zinia, Md. Abu Syed, Fahim Hasan Khan, Md. Mahbubur Rahman
TL;DR
The paper addresses the paucity of realistic Bangladeshi traffic data for autonomous driving by introducing ANNA, a monocular, smartphone-camera dataset collected from busy Bangladeshi roads. It outlines a six-step pipeline from route planning and data capture to annotation, training a YOLOv5 detector in PyTorch and evaluating with IOU-based metrics. The results show that a detector trained on ANNA markedly outperforms COCO on Bangladeshi traffic, with a MAP of $93.85\%$ and higher per-class AP, including for locally common vehicles like rickshaws and CNGs. This work demonstrates the value of region-specific data for robust camera-based AV perception and motivates expanding the dataset with more scenarios and occlusions for practical deployment.
Abstract
Recent breakthroughs in artificial intelligence offer tremendous promise for the development of self-driving applications. Deep Neural Networks, in particular, are being utilized to support the operation of semi-autonomous cars through object identification and semantic segmentation. To assess the inadequacy of the current dataset in the context of autonomous and semi-autonomous cars, we created a new dataset named ANNA. This study discusses a custom-built dataset that includes some unidentified vehicles in the perspective of Bangladesh, which are not included in the existing dataset. A dataset validity check was performed by evaluating models using the Intersection Over Union (IOU) metric. The results demonstrated that the model trained on our custom dataset was more precise and efficient than the models trained on the KITTI or COCO dataset concerning Bangladeshi traffic. The research presented in this paper also emphasizes the importance of developing accurate and efficient object detection algorithms for the advancement of autonomous vehicles.