面向车规级芯片的对象检测模型优化方法

卷积神经网络复杂的网络结构使得模型计算复杂度高，限制了其在自动驾驶等实际终端场景中的应用。针对终端场景下的计算资源受限的问题，本文从轻量化深度模型设计和车规级芯片模型部署验证两方面进行研究。针对深度模型计算效率和检测精度的矛盾，本文设计了基于中心卷积的轻量化对象检测模型，实现功耗低且精度高的模型性能。进一步，本文基于量化感知训练的模型加速部署方法在车规级芯片上开展了系统级部署验证，在车规级芯片tda4上成功实现了高效的对象检测模型，在自动驾驶场景中取得了良好的性能。

Object detection model optimization method for car-level chips

Convolutional neural networks have achieved great success in visual perception tasks. Its complex network structure makes the model computationally complex, which limits its application in actual terminal scenarios such as autonomous driving. Aiming at the problem of limited computing resources in terminal scenarios, in this paper, we conduct research from two aspects: lightweight deep model design and the model deployment and verification on car-level chips. As for the contradiction between the calculation efficiency of deep models and the detection accuracy, we design a lightweight object detection model based on the center-convolution, enjoying low power consumption and high accuracy model performance. Furthermore, based on the method of quantization aware training, we carried out system-level deployment and verification on car-level chips. We successfully implemented a high-efficiency object detection model on the car-level chips, i.e. tda4, and achieved good performance in autonomous driving scenarios.