GPR multi-target recognition using texture features of time-frequency analysis and HMM

In the process of ground penetrating radar (GPR) probe, the distance variation between antenna and buried targets is reflected by the A-scan sequence changes of the echoes. So a new target recognition method is proposed based on this variation process. In feature extraction, we combine the time-frequency analysis and image texture analysis:2-D time-frequency representations of A-scans are first calculated, and then some specific image texture descriptors are utilized to construct the feature vectors. As for target recognition, the variation process is modeled using one-way non-stride continuous hidden Markov model (HMM), according to the changing distance between antenna and targets. Experiment results show that the proposed HMM method based on the variation process of A-scan sequences performs better than support vector machine, which only relies on the unordered sets of features vectors of A-scans.