| 摘要: |
| 雷达是自动驾驶的重要组成部分，通过雷达的发射和接收信号我们可以估计目标距离和速度。随着雷达的广泛使用，不同车辆安装的雷达之间会产生相互干扰，导致本底噪声升高，严重影响目标的可检测性。近年来研究者们开始用深度学习方法来抑制雷达干扰，目前常用的方法大多是基于频域或时频域，但基于深度学习的语音分离方向的研究已证明基于时域的方法相对于其他方法的优越性。然而传统的循环神经网络（Recurrent Neural Network，RNN）对长序列信号建模十分困难，因此本文使用了自注意双路径循环神经网络(Dual-Path Recurrent Neural Network with Self Attention, DPRNN-SelfAttention)，用于在深层结构中对长序列进行建模。DPRNN-SelfAttention将长序列输入拆分为更小的块，并迭代地应用块内和块间操作处理整个序列的信息，其中输入长度与每次操作中原始序列长度的平方根成正比。实验结果表明，本文方法可以抑制多个干扰，并能估计抑制干扰后的目标幅度和相位。 |
| 关键词: 自动驾驶 干扰抑制 深度学习 DPRNN-SelfAttention |
| DOI：DOI:10.3969/j.issn.1672-2337.2022.06.012 |
| 分类号:TN959.71;U463.6 |
| 基金项目: |
|
| Automotive Radar Signal Interference Mitigation Using Dual-Path RNN with Self Attention |
|
WEN Hao， GAO Yong
|
|
College of Electronics and Information Engineering， Sichuan University， Chengdu 610065， China
|
| Abstract: |
| Radar is an important part of autonomous driving. The range and speed of the target can be estimated by transmitting and receiving signals from radar. With the wide applications of radars, there is mutual interference between radars installed in different vehicles, resulting in an increase in the noise floor, which seriously affects the detectability of the target. In recent years researchers have started to use deep learning methods to suppress radar interference, most of the commonly used methods are based on frequency domain or time-frequency domain, but the research in the direction of speech separation based on deep learning has demonstrated the superiority of time-domain based methods over time-frequency domain and frequency domain based methods. However, traditional recurrent neural networks (RNN) are very difficult to model long sequence signals, so this paper uses a dual-path recurrent neural network with self-attention (DPRNN-SelfAttention) for modeling long sequences by organizing arbitrary types of RNN layers in a deep structure. DPRNN-SelfAttention takes long sequence inputs into smaller blocks and iteratively applies intra- and inter-block operations to process the information of the entire sequence, where the input length is proportional to the square root of the original sequence length in each operation. Experimental results show that the method in this paper can suppress multiple interfe-rences and can estimate the target amplitude and phase after suppressing the interferences. |
| Key words: autonomous driving interference mitigation deep learning DPRNN-SelfAttention |
