面向智能驾驶的轻量化GwcNet立体匹配算法研究

Research on lightweight GwcNet stereo matching algorithm for intelligent driving

智能驾驶技术作为新一代智能交通系统的核心之一,受到了广泛关注。其中,感知融合技术在实现智能驾驶的精准定位和环境感知中起着至关重要的作用。在感知融合中使用的立体匹配算法更是关乎智能驾驶的安全性和准确性,是智能驾驶环境感知领域中重要的技术环节。针对智能驾驶汽车在硬件方面算力不足,但对立体匹配算法实时性和精度均有较高要求的问题,文章基于GwcNet,为智能驾驶环境感知模块设计了一种轻量化的立体匹配算法。利用ACV模型替代GwcNet立体匹配算法中参数量、运算量最大的3D卷积模块,使得算法的实时性和精度得到提升。为了进一步减少网络复杂度,减小ACV模块中的网络参数,提出一个Fast-ACV模型。最后在KITTI 2015数据集中对立体匹配算法进行对比分析。结果表明,所提出的轻量化GwcNet立体匹配算法在精度和实时性上均优于GwcNet算法。

As one of the core of the new generation of intelligent transportation system,intelligent driving technology has been widely concerned.The perception fusion technology plays a crucial role in realizing the accurate positioning and environment perception of intelligent driving.The stereo matching algorithm used in perception fusion is crucial for the safety and accuracy of intelligent driving,and is an important technical link in the environment perception field of the intelligent driving.In allusion to the problems that the intelligent driving vehicles have the insufficient computing power in hardware,but have high requirements for real-time performance and accuracy of stereo matching algorithm,a lightweight stereo matching algorithm for the intelligent driving environment perception module is designed based on GwcNet.The ACV(attention concatenation volume)is used to replace the 3D convolutional module with the maximum number of parameters and computation in GwcNet stereo matching algorithm,which can improve the real-time performance and accuracy of the algorithm.In order to further reduce the network complexity and minize the network parameters in the ACV module,a Fast-ACV model is proposed.The comparison analysis of the stereo matching algorithm is conducted in the KITTI 2015 dataset.The results show that the proposed lightweight GwcNet stereo matching algorithm is superior to the GwcNet algorithm in both accuracy and real-time performance.