Overcoming Catastrophic Forgetting in Lane Detection

Abstract

With the development of autonomous driving in recent years, algorithms for lane detection emerge in endlessly. In this paper, lane detection has improved the performance of lane detection and overcome the catastrophic forgetting problem of machine learning[20]. The study of classical physics describes a static world, while physiology focuses on evolutionary systems related to the time dimension, that is, the non-stationary world. We propose ways of lifelong learning to alleviate the catastrophic forgetting of machine learning in lane marking detection, because lifelong learning can accumulate experience over time. In real and complex road conditions, lane detection needs to take into account the data imbalance in multiple scenarios, including the dynamic changes in the number and type of lane. By comparing with traditional deep learning, we introduced the most advanced lifelong learning strategies (EWC, MAS, SCP), and applied them to the advanced backbone network ENet. A large number of experimental results under the CULane datasets prove that our model significantly improves the average performance of models, which adopts the methods of lifelong learning in a variety of complex real-world scenarios. (Abstract)