基于语义分割与深度估计的行车环境实时解析

中图分类号:TP3 Real-time semantic segmentation and depth estimation method for driving environment

摘要:

道路行车环境的实时解析是智能驾驶的关键技术，尽管神经网络在实现语义分割和深度估计上能取得不错的精度，但由于模型参数多、计算量大等问题，导致难以实现实时计算。针对该问题，提出了一个轻量化、高效的特征提取模块和一个综合考虑语义信息和深度信息的特征解码模块，在一个网络中同时完成语义分割和深度估计两个任务。在CityScapes数据集中，语义分割预测结果的mIOU为65.0%、深度估计结果的误差为0.21,并且在单个GPU上推断速度达到了65FPS，满足实时性要求。

Abstract:

The real-time analysis of the road driving environment is the key technology of intelligent driving. Although the neural network can achieve good precision in semantic segmentation and depth estimation, it is difficult to realize real-time calculation due to problems such as many model parameters and large calculation amount. Aiming at this problem, this paper proposes a lightweight and efficient feature extraction module and a feature decoding module that comprehensively considers semantic information and depth information, and simultaneously performs two tasks of semantic segmentation and depth estimation in a network. In the CityScapes dataset, the mIOU of the semantic segmentation prediction result is 65.0%, the error of the depth estimation result is 0.21, and the inference speed reaches 65 FPS on a single GPU, meeting the real-time requirements.