Identification of dynamic traffic crash risk for cross-area freeways based on statistical and machine learning methods

Abstract

Freeway traffic safety should be given great attention due to the frequent and serious consequences that arise from freeway traffic crashes. With the possibility of obtaining high resolution traffic big data, traditional freeway safety promotion methods are gradually replaced by the emerging technology of active safety control based on the real-time traffic data. However, there is still a lack of cross-area pertinence towards dynamic traffic safety currently. To overcome the defects of existing dynamic traffic crash precursors identification studies and provide precise theoretical basis for freeway dynamic safety control, this research took cross-area freeways as the research object. Firstly, based on the LOS (level of service) A–F theory and the consideration of area types, six units to be evaluated were generated (two dimensions totally: saturated/unsaturated flow, urban/suburban/mountainous). Then, conditional logistic regression based on Markov chain Monte Carlo method was adopted to quantitatively evaluate the dynamic risk of cross-area freeways. Finally, 20 related traffic flow variables were extracted from five dimensions which can reflect the dynamic traffic flow characteristics, and the machine learning approaches including random forest algorithm and Bayesian logistics regression were applied for analysis and modeling; then, crash precursors were identified for each type of freeway area, and the statistical relationship between traffic flow variables and crash risk was established via the statistical models. The results show the area types and traffic conditions of freeway are significantly correlated with dynamic traffic safety, and the crash risk in the urban area/saturated flow condition is the highest, which is 29.6 times of that in the condition of suburban area/unsaturated flow. When the traffic operates at the freeway of urban/unsaturated flow, the formation of small fleets and the lane changing behavior play a dominant role influencing crash risk; when operating at urban/saturated flow, frequent acceleration or deceleration in the fleet and frequent lane changes during the formation of the fleet have great impact towards crash risk. When traffic operates at suburban/unsaturated flow, the main crash precursors are the sudden change of occupancy and volume in a short period, as well as the formation of a small fleet. When operating at suburban/saturated flow, the crash-prone variables are the sudden increase of speed and occupancy in a short period and the change of volume. When traffic operates at mountainous/unsaturated flow, the changes of speed and occupancy in a short period and the formation of small fleets are the main factors influencing crash. When operating at mountainous/saturated flow, lane changing behavior, the sudden increase of volume, and the speed mutation of some vehicles in saturated flow are the significant precursors. The results indicate every area type of freeway has a different mechanism of traffic crash. Moreover, with the consideration of freeway area types and traffic state differences, the machine learning and statistical combination models proposed in this research can identify the relationship and mechanism between dynamic traffic flow characteristics and traffic safety more comprehensively and accurately, and the identification results in this research can refer to the further dynamic crash prediction work for cross-area freeway.