Comparative Study of CNN Models for Human Action Recognition in Detection Tasks

Aiming at the performance optimization of convolutional neural networks in human action recognition tasks, this study constructs a system evaluation framework containing eight typical architectures such as VGG19 and ResNet50. Based on the standardized dataset, the model is optimized end-to-end by deep migration learning strategy, and the dynamic learning rate adjustment mechanism is proposed. The experimental results show that NASNet-Large significantly outperforms the comparison model with 91.5% validation accuracy, but its computational complexity reaches 88.9M parametric quantities, while MobileNetV2 achieves 76.6% accuracy with 3.5M parametric quantities, demonstrating the optimal computational efficiency. Quantitative analysis reveals a nonlinear relationship between model depth and feature extraction capability: the deep network excels in spatio-temporal feature fusion, while the lightweight architecture has an obvious advantage in real-time metrics. The comprehensive score function proposed in this study effectively balances the trade-off between accuracy and computational efficiency, providing a quantitative decision basis for model selection in industrial-grade applications.