Reducing Data Costs- Transfer Learning Based Traffic Sign Classification Approach

Traffic sign detection is a very important input to vehicle navigation. To achieve both urban and highway maneuvering, driverless cars must learn to obey traffic signs, while driving and while stopping at intersections. However, in the United States, high quality public data on road traffic signs is not readily available to train self-driving cars and to recognize traffic signs such as stop, crosswalk and many more. Currently, most of the autonomous car companies keep their data proprietary. These Datasets take a huge amount of resources to collect large enough datasets for deep learning. An alternative solution is to adopt transfer learning, by exploring the use of a Residual deep neural Networks (ResNet) and employing transfer learning to detect and classify traffic signs. In the current research study, a high-quality training data is collected using NCAT's Autonomous-driving car sponsored by GM and SAE for the Auto drive Challenge Competition. The goal is to develop a deep learning traffic sign classifier model with transfer learning to save developmental resources of time, money, and labor.