Abstract: In order to improve the effectiveness of highway traffic accident detection, according to the changing characteristics of upstream and downstream traffic flow parameters at the occurrence of traffic accidents, this paper constructs a relatively comprehensive set of initial feature variables for traffic accident detection, and filters out important feature variables by using the Random Forest-Recursive Feature Elimination with Cross Validation (RF-RFECV) algorithm. The long and short-term memory (LSTM) network is trained by using significant feature variables as the input, and the hyperparameters of the LSTM network are optimized by a Bayesian optimization algorithm (BOA). Finally, real highway data are used for validation and comparative analysis, and Borderline-SMOTE is used to solve the imbalance of the traffic dataset. The experimental results show that selecting the important feature variables that are more sensitive to traffic accident detection can improve the detection accuracy, and the detection effect of LSTM is significantly better than that of random forest (RF) and support vector machine (SVM).