Abstract
Abstract With the increasing demand for intelligent sports training, traditional motion recognition methods based on visual or sensor data face challenges such as strong environmental dependence, high computational complexity, and insufficient interpretability. This paper proposes an innovative deep neural decision forest model, which builds an end-to-end motion analysis system by fusing the feature extraction ability of convolutional neural network and the hierarchical classification mechanism of decision forest. The model uses multi-granularity scanning technology to capture multi-scale features of time series data, the differentiable decision tree is used to realize the hierarchical classification of rule interpretation, and the decision path and feature importance weight are visualized to provide the coaches with “white box” decision-making basis. The experimental results show that the decision path consistency of this model reaches 94.6%, which is significantly better than the black box CNN model. The experimental results show that on the three public data sets of WISDM, UCI-HAR and DAPHNet, the model action recognition accuracy rate reaches 98.5%, and the F1 score is 98.3%, which is 4.3 percentage points higher than that of traditional convolutional neural networks. In a noisy environment (signal-to-noise ratio of 10 dB), the recognition accuracy is still 92.7%, and the inference time of embedded devices is only 12.1 ms. This study confirms the effectiveness of deep neural decision forest in complex sports scenarios. Its innovation is reflected in transforming non-differentiable decision trees into differentiable forms to achieve end-to-end training, developing multi-granularity feature extraction mechanisms to enhance time series modeling capabilities, establishing an adaptive integration framework to improve system stability, and providing a new technical path and theoretical support for precise sports training. The core innovations of this paper can be condensed into three points: (1) structure differentiability: this paper transforms the non-differentiable decision tree into differentiable form through soft binning function and Kronecker product for the first time, and realizes the end-to-end joint optimization of CNN feature extraction and decision forest classification; (2) time series multi-granularity: this paper proposes a sliding window multi-granularity scanning mechanism to synchronously capture the transient impact and long-term rhythm of motion data; (3) dynamic integration: this paper designs an adaptive weight allocation strategy based on the accuracy of validation set, which replaces the static voting mechanism of traditional random forest, and significantly improves the robustness in noisy environment. The system can not only output action labels in real time, but also generate action quality scores (0–100 points) based on the predicted probability entropy value, and automatically correlate the key feature deviations in the decision path. In the field test, the system pushes the visual feedback report to the coaches through the tablet computer with a delay of 12.1ms, which supports the formulation of personalized correction scheme, and realized the closed loop from ‘action recognition’ to ‘intelligent coaching’.