Road accidents are one of the most predominant factors for deaths throughout the world. With the inclusion of several driver assistance systems, intelligent vehicles are becoming peak of the automotive industry to mitigate accidents. Autonomous vehicles are widely considered safer, because of the introduction of advanced robotics and Advanced Driver Assistance Systems (ADAS) into the task of driving. However, the main challenge for AVs is to properly detect dangerous situations and react properly to avoid potential collisions. To overcome this challenge, it is important to assess current traffic situation and vehicle dynamics for real-time collision prediction. This thesis provides an insight to identify and predict dangerous driving behaviour for autonomous vehicles in an uncontrolled intersection for rear-end collision scenarios. A large naturalistic driving dataset containing single vehicle data of position, speed and heading is analyzed to predict future conflicts by utilizing machine learning classification techniques.
To that aim, vehicle level data are collected using sensors installed on a vehicle, which deliberately passes through an uncontrolled T-intersection. The vehicle passed approximately ten times in each of the six possible manoeuvres. A circular area of interest with radius of 35 meters is selected around the center of intersection. Based on this bounding area, vehicle trajectories are extracted from position data based on their entry and exit points. Trajectories are then time-shifted, so as to imitate interactions among them and develop rear-end collision scenarios. Finally, Time-to-Collision (TTC) is used as a surrogate safety indicator to identify dangerous behaviour.
A total of 11,208 gap observations are counted in all six manoeuvres in the bounding area. Among them 35.96% observations are marked as dangerous, where TTC lie below the threshold value of 1.5 seconds. It is observed that TTC gets lower when the vehicle approaches to the intersection. Moreover, there is an inverse relationship between TTC and speed difference. High difference of speed between the following vehicle and lead vehicle leads to lower TTC and results in dangerous situation. On the contrary, low speed difference shows high TTC and low collision risk. It is observed that TTC decreases exponentially with increase in speed difference between the following and lead vehicle.
Finally, different machine learning classifiers are tested to classify and predict dangerous situations considering speed difference as the independent variable or predictor. After analyzing performance matrices, it is observed that Random Forest (RF) performs better than other classifiers in terms of different performance matrices and gives a lower rate of false alarm (less than 7%). Area under the ROC curve also increases for RF. Later on, RF classifier is employed in all the six manoeuvres to classify dangerous driving behaviour. However, in some manoeuvres, it gives higher false prediction due to the high imbalance between safe and collision-prone test samples. It is expected that more sophisticated real-world traffic data and integration of more advanced classification techniques like imbalanced learning or deep learning are more likely to give better prediction of collisions.
«
Road accidents are one of the most predominant factors for deaths throughout the world. With the inclusion of several driver assistance systems, intelligent vehicles are becoming peak of the automotive industry to mitigate accidents. Autonomous vehicles are widely considered safer, because of the introduction of advanced robotics and Advanced Driver Assistance Systems (ADAS) into the task of driving. However, the main challenge for AVs is to properly detect dangerous situations and react properl...
»
Login
BibTeX