Stroke patients often exhibit hemiparetic gait, and visual gait analysis can be applied to detect such changes. However, publicly available pathological gait datasets are small in scale and lack detailed grading of hemiplegia severity. Furthermore, state-of-the-art deep learning algorithms for gait analysis usually have a high need for parameter size and computational complexity, leading to low performance on small-scale pathological gait datasets. To address these challenges, this study designs a lightweight hemiplegic gait recognition system. The system utilizes an attention-based lightweight convolutional neural network (CNN) to access hemiplegic gait performance. By linear splicing grouped convolution at different scales, high-efficiency features can be obtained at a low cost. Additionally, a multidimensional hybrid lightweight attention module is introduced to assist CNN in focusing on distinctive features in both spatial and channel dimensions, achieving a good balance between system effectiveness and lightweight design. Moreover, a hemiplegic simulation gait dataset is constructed, specifically for hemiplegic gait recognition to support model training and testing. The results demonstrate that the proposed network that uses only 1/53 parameters of VGG-19 improves the accuracy of gait recognition to 96.91%, which is higher than that of pre-train VGG-19. Compared with other lightweight SOTA methods, it also has the advantage of accuracy. The system has low development costs and can be deployed on mobile devices. It supports real-time detection, providing a feasible solution for home-based pathological gait analysis.