Abstract: The accurate classification of targets concealed in non-line-of-sight （NLOS） regions， such as armed personnel or unmanned aerial vehicles behind building corners， holds significant value for modern urban warfare and intelligent driving systems. However， the complexity of electromagnetic wave propagation in NLOS scenarios often introduces multipath interference， blurred path boundaries， and signal distortion in echoes， thereby undermining the reliability of traditional classification methods that rely on geometric modeling or handcrafted features. To address these challenges， this study proposes a multipath feature fusion network for NLOS target classification. First， the echo characteristics of four representative target types were analyzed： pedestrians， soldiers carrying anti-tank weapons， quadrotor drones， and tank models. Given the variability in multipath returns and boundary blurring， an adaptive multi-scale feature extraction module was designed. Specifically， small-scale convolutional kernels were employed to capture fine-grained， high-amplitude features in single-path echoes， whereas large-scale kernels were used to extract global structural information related to multiple propagation paths. Additionally， an attention mechanism was incorporated to effectively fuse salient multipath features with local scattering information， thereby enhancing classification accuracy under NLOS conditions. Finally， real-world measurements collected with a multipath radar sensing system were used for validation. The experimental results demonstrate that the proposed network achieved a target classification accuracy of 99.6%. Comparative experiments confirmed the effectiveness and robustness of the proposed method.