Developing vehicle-based traffic conflict prediction algorithms using deep learning

With the rapid growth of artificial intelligence technologies such as big data analytic, machine learning, and image recognition, the vehicle industry has undergone dramatic changes. The vehicle is no longer a simple mechanical structure, but it engages with the driver, environment and infrastructure. For instance, intelligent vehicles aim to improve vehicle and driver safety by utilising multiple Advanced Driver Assistance Systems (ADAS). These emerging technologies in the automotive industry have introduced safety-related challenges and consequently, research has attempted to address these challenges by designing and developing proactive safety management systems for these vehicles. In particular, to proactively mitigate the risk of collision, there is a need to predict traffic conflicts to prevent collisions.

Existing safety prediction algorithms assess and quantify the threat level surrounding the ego-vehicle. However, they are not able to plan the best response to a fully unexpected dangerous situation while driving. Therefore, it is important that the algorithm has the ability to cope with uncertainties since not all situations are ‘car-following’. Previous research has not taken this uncertainty into account, so it is desirable to develop robust algorithms which are not restricted by the predefined movement patterns of the vehicle. In fact, the readily available safety algorithms estimate the threat level based only on one factor: Time-To-Collision (TTC). However, using only a single factor with a pre-define threshold for every traffic situation can be limiting. This is because it cannot handle all scenarios and ignores all uncertainties while completely disregarding other factors that influence traffic conflicts such as speed variance, traffic density, speed and weather conditions. Considering all these factors in detecting traffic conflicts is rather complex as it requires an integration and mining of heterodox data, the unavailability of traffic conflicts and conflict prediction models capable of extracting meaningful and accurate information in a timely manner. In addition, the algorithm has to effectively handle large imbalanced data since the majority of the cases are safe traffic dynamics with traffic conflicts being the minority. Such complexity in the imbalanced conflicts dataset could create biased and inaccurate predictions. Moreover, testing and validating the developed traffic conflict algorithm would be the key to prove their effectiveness. However, the complexity of modelling all possible combinations of traffic situations and environmental conditions make this approach challenging.

To overcome these limitations, this thesis presented a centralised digital architecture to capture all the data and developed a methodological framework to find the optimal machine learning technique to implement within the algorithm, by making use of cost-sensitive learning to alleviate the detrimental effects brought by class imbalance. Five machine learning techniques were modified and utilised including: logistic regression, support vector machines, Deep Neural Networks (DNN), Long Short-Term Memory (LSTM) and a hybrid method of LSTMs with Convolutional Neural Network (LSTM-CNN). Unlike many existing studies, this study employs many interconnected factors to reliably predict a traffic conflict in real-time. Comparative analyses were also undertaken by incorporating varying input factors to each system (e.g., surrogate safety measures (SSM), traffic, weather, vehicle-related data and hybrid of factors) to determine which factors are important and which increase sensitivity to traffic conflict predictions. Additionally, a novel addition to the algorithm was to estimate the factors by considering uncertainty, i.e., accommodating for more stochastic movement patterns of vehicles.

To train the machine learning technique and identify patterns of what arises from the factors to result in a traffic conflict, both the output (i.e., tracked vehicles which are classified as ‘threats’) and the input (i.e., corresponding estimated factors) were required. However, traffic conflicts data was not known a-priori and an automated video analysis method was developed by which traffic conflicts were identified using a faster Regional–Convolution Neural Network (R-CNN). This data is then integrated with highly disaggregated microscopic traffic data and in-vehicle sensors data collected from a section of the UK M1 motorway using an instrumented vehicle. It was found that DNNs outperform other techniques in predict conflicts. Such promising results reflect that DNNs can be further applied to deepen our understanding in predicting traffic conflicts in order to design more reliable primary safety systems for intelligent vehicles. Results also show that by adding multiple factors a significant difference (64.5%) in sensitivity is observed than when adopting TTC only for a 10% false alarm rate. Additionally, when considering uncertainty, a consistently higher sensitivity value and a 5% increase in AUC value was observed when compared to traditional estimation. This extends the systems’ application for a wider spectrum of traffic scenarios. To validate the safety performance and prediction accuracy of the developed algorithm, an integrated simulation framework was developed. The framework consisted of a submicroscopic simulator, which provided an appropriate testbed to develop a scenario to test the effectiveness of the algorithm, and a microscopic traffic simulation tool to simulate the surrounding traffic accurately based on real-time data. Rear-end and lane change traffic conflict scenarios were developed. The validation results from the integrated simulation framework are significant. In fact, approximately 80% of rear-end conflicts and 73% of lane change conflicts were predicted by algorithm for a 10% false alarm rate. Despite that the algorithm was not trained using the virtual data, the sensitivity is high. This highlights the transferability of the algorithm to similar road networks. Consequently, this algorithm has the potential to be used in intelligent vehicle as an ADAS for improving traffic safety, presenting a viable solution for implementation within connected and autonomous vehicles.