Optimization of Lightweight Convolutional Neural Networks in Real-Time Motion Posture Recognition

In view of the problems of poor real-time performance and difficult mobile terminal deployment caused by high network complexity of existing motion posture recognition models, this paper proposes a lightweight convolutional neural network optimization method to achieve high-precision real-time recognition. First, a lightweight backbone network based on deep separable convolution is designed (MobileNetV3 improved architecture), and the number of parameters is compressed to 0.78M through channel pruning. Secondly, the Convolutional Block Attention Module (CBAM) is introduced to enhance the extraction of key joint features, and the adaptive feature fusion technology is used to splice shallow details with deep semantic information. Finally, a hybrid quantization strategy (weight 8bit + activation value 4bit) is deployed to reduce the model size to 1.2MB. Experiments on the COCO- WholeBody dataset show that the optimized model has an inference speed of 42FPS on the mobile terminal (Snapdragon 865), which is 3.7 times higher than the baseline model ResNet-50, and the maximum average precision (AP) is 82%. This verifies the effectiveness of this method in real-time motion monitoring of wearable devices.