Electrically Large Complex Objects Recognition Based on Gated Recurrent Residual Network (GRRNet)

In this paper, a novel deep model based on gated recurrent residual network for electrically large complex objects recognition is proposed. It can fully exploit the data envelope information and temporal correlation to improve the system recognition performance. Electromagnetic (EM) scattering property measurements for electrically large objects are costly and time-consuming, affected by various environmental factors. The high-frequency approximate technique, namely the shooting and bouncing ray method (SBR), is introduced to quickly acquire high resolution one-dimensional range profile (HRRP) of electrically large complex objects. Both the corner reflector and the model car are measured to validate the accuracy of the SBR method. The method is employed to establish HRRP database for various vehicles in traffic scenarios. Deep learning can automatically study data deep features and show outstanding performance in various classification tasks. The residual network (ResNet) and gated recurrent unit (GRU) models are combined to capture and aggregate scattering information of objects. ResNet uses 1-D convolutional kernels and residual blocks to efficiently capture the scattering information within each distance cell while avoiding gradient vanishing or gradient explosion issue. GRU aggregates scattering information along the spatial dimension to construct object feature representations. The combination of them can take advantage of their respective strengths to fully mine the information of HRRPs. Compared with the conventional methods, the features extracted by the proposed model from each class are more concentrated shown in the result of t-distributed stochastic neighbor embedding. The deep model exhibits a superior average recognition rate up to 95.56%, significantly higher than existing methods. It shows robustness to noise, thereby showcasing good potential for practical applications within the Internet of Vehicles (IoVs).