Figures (4)  Tables (2)

    • Figure 1. 

      Keyword clustering network diagram from the English literature studied.

    • Figure 2. 

      Schematic diagram of the kernel density estimation principle[29−32].

    • Figure 3. 

      Schematic diagram of the neural network model.

    • Figure 4. 

      Relationship diagram of research scope.

    • Research perspectiveAccident identification methodIndexApplicable conditionsAdvantageDisadvantage
      Micro levelAccident frequency methodNumber of accidentsSuitable for a specific range of smaller intersections or streets, etc.Single indicator;
      Easy to determine;
      With good operability      		The judgment results are highly subjective, and the differences in traffic conditions and road environment are not considered;
      High error rate
      Accident rate methodAccident rateSuitable for longer sections or regional roadsComprehensive consideration;
      Strong operability and intuitive results      		Lack of consideration of the severity of the accident;
      A gap with actual traffic conditions
      Empirical model methodPredicted number of accidentsSuitable for level intersections of roads.The model is intuitive;
      It can solve the problem of random fluctuation      		Difficulty in collecting a large number of typical level intersection statistics;
      Cannot be used for road segment or area identification
      Equivalent accident number methodEquivalent AccidentsSuitable for roads with similar road conditions and stable traffic volumeConsidering the differences in the severity of accidents;
      The evaluation results are objective and comprehensive      		Factors such as traffic volume and road conditions are not considered
      The choice of weights is subjective
      Traffic conflict lawsDistance, speed, timeSuitable for urban roads or specific road sections and intersectionsThe dependence on accident statistics is small;
      The theoretical basis is sufficient; The cycle is short      		Conflict investigation workload is heavy;
      Difficult to model, poor portability
      The influence of road environment, etc. is not considered
      Meso levelQuality control methodUpper and lower bounds on the combined accident rateSuitable for road networks or road sections with roughly the same traffic conditionsThe calculation is simple, the theory is perfect, and the scope of application is wide, considering the random fluctuation problemThe workload is large and accurate traffic data is required;
      Does not consider spatiotemporal accumulation and dynamic patterns of spatiotemporal evolution;
      Confidence selection is subjective
      Empirical BayesBayesian prior and posterior estimates, recognition thresholdsSuitable for intersections with roughly the same road and traffic conditionsHomogeneous road data is considered to avoid the influence of the regression effect caused by the randomness of accidents, and the prediction accuracy is highExcessive requirements for the completeness of historical data
      Does not consider spatiotemporal accumulation and dynamic patterns of spatiotemporal evolution
      The calculation process is complicated
      Macro levelGrey evaluation methodInfluencing factors and indicators constitute an evaluation setSuitable for regional road networkClear meaning, clear algorithm, strong practicabilityHave a certain degree of subjectivity;
      Data indicators are too single;
      Evaluation accuracy is low
      Cumulative frequency methodNumber of accidents per kilometer, accident rate per vehicle kilometerSuitable for regional roads with poor traffic conditions and different accident conditionsWide application, relatively mature judgment threshold, high practical valueThe selection of unit length has a great influence on the results;
      Factors such as traffic volume and accident severity are not considered;
      There may be peak clipping
      Regression analysisRegression Model Predicts Number of AccidentsSuitable for regional road networks and roadsConsidering factors are comprehensive and the scope of application is wideThe algorithm is relatively simple;
      The results are subjective and difficult to apply in practice
      Cluster analysisEquivalent total Accident rate value, accident-prone point threshold valueApplies to the entire road network or to specific roadsThe results are reasonable and accurate, and the positions that are easily missed by the traditional method can be identifiedThe number of indicators used is limited and the accuracy is not high
      The results have certain limitations

      Table 1. 

      Comparison of identification methods for accident-prone points.

    • PastNow
      Single typical accident dataAccident data, road attributes, spatial attributes
      Specific road intersectionSpatiotemporal road network
      Road segments by distanceDivide homogeneous road segments
      Simple historical accident statisticsSpatiotemporal accident data
      Traffic accident recordTraffic travel information
      Number of traffic accidents or conflictsImage, heat map recognition

      Table 2. 

      Comparison of identification data.