In vehicular communications using IEEE 802.11p, estimating channel frequency response (CFR) is a remarkably challenging task. The challenge for channel estimation (CE) lies in tracking variations of CFR due to the extremely fast time-varying characteristic of channel and low density pilot. To tackle such problem, inspired by image super-resolution (ISR) techniques, a deep learning-based temporal spectral channel network (TS-ChannelNet) is proposed. Following the process of ISR, an average decision-directed estimation with time truncation (ADD-TT) is first presented to extend pilot values into tentative CFR, thus tracking coarsely variations. Then, to make tentative CFR values accurate, a super resolution convolutional long short-term memory (SR-ConvLSTM) is utilized to track channel extreme variations by extracting sufficiently temporal spectral correlation of data symbols. Three representative vehicular environments are investigated to demonstrate the performance of our proposed TS-ChannelNet in terms of normalized mean square error (NMSE) and bit error rate (BER). The proposed method has an evident performance gain over existing methods, reaching about 84.5% improvements at some high signal-noise-ratio (SNR) regions.

在使用IEEE 802.11p的车辆通信中，估计信道频率响应（CFR）是一项极具挑战性的任务。信道估计（CE）的挑战在于，由于信道和低密度导频的极快时变特性，跟踪CFR的变化。为了解决这个问题，受图像超分辨率（ISR）技术的启发，提出了一种基于深度学习的时间频谱信道网络（TS-ChannelNet）。在ISR的过程之后，首先提出带有时间截断的平均决策导向估计（ADD-TT），以将导频值扩展到暂定CFR中，从而粗略地跟踪变化。然后，为了使准确的CFR值准确，超分辨率卷积长短期存储器（SR-ConvLSTM）用于通过提取数据符号的足够时域频谱相关性来跟踪信道极端变化。对三个代表性的汽车环境进行了研究，以证明我们提出的TS-ChannelNet在归一化均方误差（NMSE）和误码率（BER）方面的性能。与现有方法相比，该方法具有明显的性能提升，在某些高信噪比（SNR）区域达到了约84.5％的改进。