Modeling head-on crash probability on two-lane undivided highway using vision-based classified trajectory

Abstract

Head-on crash on rural undivided highways has become most inevitable event now-a-days. Thousands of people are either losing their life or accepting life-long disability in these crashes due to lack in traffic safety; whereas traffic safety aims at the reduction of fatalities causing from crashes among road users. However, this aim of traffic safety cannot be achieved without proper understanding of the crash mechanism. Traffic safety analysis requires historical records of crash data and these crash data lack in availability and quality. Furthermore, to prevent crashes by using historical records is a reactive approach requiring large amount of crash data. Hence, these problems related to crash data motivate the development of surrogate measures of safety. Surrogate safety measures are based on crash probability rather than on the observation of actual crashes. Thus, this thesis endeavors to develop a model that estimates head-on crash probability from vision based classified vehicle trajectory. The crash probability estimation model formulation considered: (1) drivers’ overtaking decision (OD); and (2) time-to-collision (TTC) on two-lane undivided highway. Drivers’ overtaking decision was modeled using nonlinear random parameter multivariate binary logistic regression. It considered variables related to both traffic (i.e. vehicle speed and spacing) and drivers’ characteristics (i.e. aggressiveness). In contrast, TTC was determined using a new formulation that considered the dynamic acceleration of the vehicles in addition to their speed and spacing. Incorporation of two new parameters, i.e. overtaking importance factor (OIF) and crash frequency parameter (CFP) enabled the estimation of crash probability combining OD and TTC. However, calibration of these models (OD and TTC) requires high frequency and well-structured vehicle trajectory data. In this regard, background subtraction technique along with Kalman filter was used to obtain vehicle trajectories from real-time video. Background subtraction technique was applied using a newly developed background estimation model. A number of theories were proposed to define different components of a video image. Specifically, first-order model for illumination variation and Fourier series for incorporating traffic arrival patterns were considered to define background and foreground, respectively. These definitions were utilized to formulate the traffic detection problem and subsequently three adaptive dynamic background models were developed to solve it. The third model, which incorporates both luminance and pollution controlling parameters addresses the traffic detection problems and limitations faced by the first and second models. This final model consists of two parameters: (1) luminance controlling parameter and (2) pollution controlling parameter. Furthermore, this study reveals newly discovered ‘ghost’ formation due to taking geometric mean of background and the subject frame, which has been termed as transparency effect. Foreground segmentation was done to get a binary image. Foreground segmentation uses a new heuristic dynamic threshold-difference ( ) function for determining per pixel threshold. The shadow of the vehicle was removed considering its physical characteristics by newly presented PNS (Positive Negative Segmentation) technique. Impulse flow waves and aggregated pictorial speed were computed after shadow removal. Impulse flow waves were eventually rectified and cumulated into actual flow. On the other hand, pictorial speed was converted into actual speed using calibration equation considering perspective error. After the foreground segmentation, connected component analysis is applied to find the geometric properties (i.e. centroid, area) of the detected object. Kalman filter was applied to get the tracking data from the detected object. This tracking data is aggregated into trajectory by means of data processing algorithm. Three different types of data were collected for this thesis work. The first one consists of six videos for calibration and validation of the background model. These videos contains a mixture of mild to hard challenges such as gradual to sudden illumination variation, stop and go traffic situation. In the second one, three different locations in Dhaka city were chosen to validate the traffic measurement mechanism. In the third one, a video (9000 sec) was captured from a two-lane undivided rural highway containing high speed uninterrupted vehicles. To avoid detailed object detection, the mounting height of the cameras was kept at 20ft and their angle was less than 45 degrees in each of the cases. Variable inputs required for calibrating the OD model were generated by constructing adjacency matrices among the detected vehicles from the third video. Analysis over these inputs shows that, lower front vehicle speed invokes the overtaking maneuver. Moreover, the bus and car drivers are found to be more aggressive drivers while overtaking. Exploiting these inputs, Metropolis-Hastings algorithm was applied to obtain calibrated parameters of the OD model for different classes of vehicle. Calibration result shows that subject vehicle speed and the subject-opposing spacing are the most significant variables influencing the overtaking decision on two-lane undivided highway. In another way, lead vehicle speed and subject vehicle aggressiveness also influences the overtaking decision largely. Opposing vehicle speed found to be least influencing in making overtaking decision. Besides, the maximum head-on crash probability for different types of vehicles while completing overtaking maneuver were determined and it was found that bus has the largest one. Finally, the nomographs established in this thesis ensures easy determination of the crash probability.