<p>Recognizing emotions from gait is an emerging area of biometric research, which has applications in medicine, robotics, smart home design, risk assessment, and rescue operations. This article presents a novel deep learning architecture for gait emotion recognition. It utilizes a fusion of latent deep features, extracted from an efficient sequential neural network, with discriminative domain-specific handcrafted features. The results from the numerous experiments used to fine-tune the proposed bi-modal deep neural network (BMDNN) are provided. Additionally, the importance of each component of the architecture is validated via an ablation study and the effects of various Laban Movement Analysis (LMA) groups on the performance of the proposed architecture are analyzed. The proposed architecture outperforms state-of-the-art methods in all emotional classes on two benchmark datasets.</p>