Objectives  Traffic congestion detection is one of the key points and difficulties of urban traffic management. The existing congestion detection methods are based on road sections, which is not conducive to the extraction of spatiotemporal evolution information of congestion. Moreover, most of the detection only involves the degree of congestion but lacks the congestion type identification. With the classification and regression tree (CART) algorithm, this paper proposes a method for the classification and detection of traffic congestion points, which takes the road section point as the detection unit. In the practical application of this method, congestion points and their categories can be detected in real time according to the average running speed on the road section.

  Methods  Firstly, the road section is divided at a specific interval and mapped to be road section points. According to the abnormal rules and patterns of spatiotemporal road conditions, the spatiotemporal evolution patterns of four congestion types are analyzed with road section points as units. Secondly, the spatial and temporal sequence of road conditions of road section points is extracted on the basis of the road condition detection of road sections, and the spatial and temporal sequence of road conditions is classified and labeled according to different congestion types. Thirdly, four speed indexes are selected to constitute the attribute set of samples, and the speed of each road section point at each peri‍od is extracted according to the attribute set, which forms the dataset of decision tree learning. Finally, with the CART algorithm, the optimal model is obtained by the training with cross-validation to achieve the best generalization ability.

  Results  This paper proposes a classification and detection method of traffic congestion points based on CART. On one hand, congestion point detection is added to refine the basic unit of congestion detection, and on the other hand, congestion point type detection is also involved. Classification and detection of congestion points is helpful to improve the efficiency of traffic management.

  Conclusions  The proposed method is compared with the support vector machine classification model, and the experimental results show that the method in this paper has higher accuracy, higher recall rate, and better classification and detection timeliness.