Vehicle detection on unstructured roads based on Transfer learning

Deep learning has seen tremendous progress in various fields. Autonomous Vehicle is one such field wherein the central branch of Computer Vision i.e Object detection serves as a key task. However, the performance of various algorithms depends on the scenarios where they are used. In this paper, an attempt is made to test the algorithms on unstructured roads with diverse data distributions. For this purpose, the single-stage network, YOLOv2 is demonstrated for vehicle detection on which transfer learning is applied. Along with the base model, the network uses Berkeley Deep Drive (BDD) dataset for knowledge transfer. Using BDD with YOLOv2 network, the model gives mAP of 72.4% and when the same model is tested in an unstructured environment using Indian Driving Dataset (IDD), it gives mAP of 56.76%.