Model Distillation for Lane Detection on Car-Level Chips

As one of the most essential perception modules in the autonomous driving system, the lane detection module must accurately and efficiently detect each traffic lane’s location, color and type to secure the vehicle’s safety. State-of-the-art lane detection models have already shown great performance using deep and complex architecture. However, since the embedded system of car-level chips such as TDA4 supports limited computing resources and operators, these models could not be deployed directly. Besides, simply compressing models using a smaller backbone usually yields unsatisfied performance. In this work, to tackle these problems, we propose a lightweight and practical lane detection model based on model distillation. Specifically, our model learns the knowledge of a pre-trained teacher model using spatial softmax. We also re-train our model using the method of quantization aware training to further compress the size of model. Compared with the original PINet [1] model, experimental results on TDA4 demonstrate that our proposed method merely consumes a quarter of the parameters while enjoying a close detection precision of 0.7524 and an IoU score of 0.7621.