基于DRHMM的高速公路场景车辆换道意图识别方法OA

To address the issues of poor generalization,susceptibility to catastrophic forgetting and degraded recognition performance due to imbalanced sample distribution in traditional vehicle lane-changing intention vecognition models under dynamic high-speed driving scenarios,the dynamic replay hidden Markov model(DRHMM)integrating experience replay strategy was proposed.Based on the NGSIM I-80 highway dataset,three types of valid trajectories including straight driving,left lane-changing and right lane-changing were extracted through trajectory splicing and reconstruction.Using Gaussian distribution parameter fitting,the lateral displacement and lateral velocity were selected as input features for the model.The experience replay strategy was introduced into the Gaussian mixture model-hidden Markov model(GMM-HMM)framework,and the experience replay buffer was constructed to store historical training samples.By mixing the new and old samples during training,the catastrophic model forgetting was suppressed,and the adaptability to dynamic driving environments was enhanced.The sliding time window of 3.00 s was set to conduct the comparative experiments among three models of HMM,GMM-HMM and DRHMM.The results show that by the DRHMM,the lane-changing intention recognition accuracy of 92.99％is achieved with recognition delay of only 0.69 s,which demonstrates significantly better overall performance than the comparison models.