Abstract：To solve the problems of low recognition accuracy and slow convergence speed for driving behavior in driving assistant technology, a new driving behavior recognition method was proposed based on improved LRCN model. The self driving behavior data sets were adopted as input samples and were processed by Pyramid down sampling and Gauss mixture model feature extraction preprocessing algorithm, and the standard video image sequence was obtained. The image sequence was introduced into the model based on convolutional neural network and gating unit recursive network, and the optimization was conducted to get the final result of convergence. The model was calculated on the GPU with Keras framework, and environment adaptability, preprocessing algorithm and model comparison experiments were carried out respectively. The results show that the pretreatment algorithm can guarantee the convergence of the proposed model and can improve the robustness of the model recognition in different scenes and different test objects. The average recognition accuracy in the self building data set reaches 94.3% and is 4.7% higher than that of the traditional LRCN model. The model also has faster convergent speed and stronger generalization ability.
2018, Vol. 39 
