The traditional basketball serving training method mainly relies on experience and has a strong subjectivity, which affects the improvement of athletes’ technical level. Therefore, this study proposes a serving behavior analysis method based on improved convolutional neural networks and field programmable gate arrays. By introducing non-maximum suppression, adaptive thresholding, and RMSProp optimizer to enhance the edge detection module, and deploying the module on a field programmable gate array to achieve hardware acceleration. Conduct experiments based on three publicly available datasets, KTH, CelabA, and RAF-DB, as well as 250 custom basketball serving behavior datasets. The results confirmed that the image processing algorithm began to converge after 70 iterations, with a final convergence loss value of 0.1 and an average feature recognition accuracy of 96.5%. Compared with the traditional convolutional neural network’s 94.1%, it is improved by 2.4%. The real-time processing performance of the proposed method reaches 42.8 frames per second, which is about 3.5 times that of traditional convolutional neural networks, indicating its high processing efficiency. The experimental data revealed that the ratio of the athlete’s serve height to athlete’s height within [1.35 ± 0.02, 1.42 ± 0.03] is the optimal serve height. The basketball serve behavior recognition model designed in this study has higher accuracy and has certain application potential in the field of behavior analysis.