Enhancing an unbalanced data-driven recognition model for identifying dangerous car-following behavior during truck movement interruptions
In this study,we aim to detect dangerous car-following behavior of passenger cars during truck movement interruptions on two-lane highways in mountainous areas.To achieve this,we utilize drone video and video trajectory extraction technology to extract vehicle trajectories.Additionally,we employ the Synthetic Minority Oversampling Technique(SMOTE)to address the issue of unbalanced trajectory data by oversampling.Furthermore,we leverage the K-means algorithm to cluster different driving behaviors,the car-following behavior is classified into two categories:dangerous and safe.To assess the risk,we select three triggers of dangerous car-following behavior:urgent following behavior,excessive offset,and speed variation behavior.We use three measures-countdown to Time-To-Collision,relative lateral offset,and velocity variation coefficient-as the Measure of Driving Risk(MOR).These MOR values,along with clustering calibration labels,are utilized as input variables for the recognition model.We establish a recognition model for identifying risky car-following behavior using the Light Gradient Boosting Machine(LGBM).Additionally,we validate the effectiveness of the model using other algorithms such as the Support Vector Machine(SVM),Random Forest(RF),and Adaptive Boosting(AdaBoost).Using a mountainous two-lane highway in Yunnan province as a case study,a total of 543 pairs of minibus trajectory data were extracted alongside truck data.Following data preprocessing,467 pairs of effective follow-up data were identified.After sampling processing and clustering calibration,the results revealed that over 30％of the passenger cars exhibited risky follow-up behavior when following trucks.Additionally,the accuracy rates for the straight and curved road recognition models for risky follow-up behavior were 95.49％and 95.48％,with recall rates of 96.93％and 95％,and F1 values of 96.20％and 95.24％,respectively.The performance of LGBM in both curved and straight road segments is more stable compared to SVM and RF,which exhibit poor stability and low accuracy rates.The LGBM-based dangerous car-following behavior recognition model demonstrates high accuracy,stability,and promising applications in cooperative vehicle infrastructure systems,automatic driving,etc.
NETL
NSTL
万方数据
