A system design for determining traffic accident risk from real-time video images

Observation of behaviors that endanger the safety of life and goods in traffic, understanding its causes and take the necessary precautions associated with it is one of the important research topics nowadays. Provision of traffic flow on a regular basis may be possible with controlling driver behaviors by improved deterrent precautions. Provision of traffic flow also brings psychological, temporal and material gain for all the elements that make up traffic. However, perhaps the greatest yield of this condition is undoubtedly related to the protection of human life. One of the main goals of this study is to facilitate the control of the traffic flow and scheme by providing an automated analysis of the vehicle behaviors.In this study, a system is designed and implemented for examining the motion trajectories of the vehicles that are tracked by using the traffic videos, learning road models of these trajectories lead to traffic camera field of view, detecting abnormal behaviors of the vehicles with the help of road models and also estimating the probabilities of accidents according to relations with each other. For this purpose, firstly, behaviors of the vehicles that are moving on traffic camera viewing angle are learned. After application of various pre-processing methods to the video frames, foreground segmented from the background by using Adaptive Gaussian Mixture Model. Then, the central point of all the vehicles are detected and tracked by using Lucas-Kanade optical flow method. The road is modelled by using trajectories of vehicles that are produced after tracking process. For this purpose after clustering all motion trajectories, common road models that are representing each cluster and expressed as Hidden Markov Model with Gaussian Mixture are found. In the implementation phase of the system, the behavior of the vehicles is examined according to produced common road models and, according to the result of these examining, anomaly traffic conditions are easily detected. Therefore, angle and distance information between partial trajectories of vehicles that are tracking, and common road models, are used for classification process. Following this process, an accident risk factor value is obtained in consequence of assessment of current speed, position, movement direction parameters of vehicles and detected anomalies of vehicles. Then accident risk faktor is evaluated in three categories consisting of low, medium and high risk level. According to the variation of accident risk factor over time, an alarm indicating the risk of an accident is generated. As a result of learning common road models, working performance of the system is increased by examining accident risk factor only in the road models that an accident can possibly happen.According to experimental results within the scope of this study, %80 accuracy and %68 precision rates were obtained for determining traffic accident risk. The tests concluded that accidents can be estimated 2.9 to 1 second before. It?s observed that the performance of the system can be changed according to camera viewing angle, traffic density, correctly detection and tracking of moving vehicles.As a result of the implementation of the proposed process, operator-dependent structure can be replaced by a structure that is more flexible and improved with an automatic analysis of the elements that make the traffic and production of risk factors in the traffic co-ordination centers. One of the most important features of this model is that road model is automatically trainable. In this way, by using traffic cameras installed presently, it will be possible to focus on the relevant points when only abnormal situations and the risk of accidents are formed. This helps the camera network to become dependent on minimum human resources and also will help to extend. In addition, after the identification of situations that create anomalies and prediction of accidents, dispatching police teams and medical teams to the region may accelerate the process. Currently, in many metropolitans, installed traffic cameras are a great advantage in terms of the applicability of this system.