Big data processing framework for pedestrian accident risk analysis and prediction using computer vision

Abstract

and (2) requiring the high cost and time-consuming to obtain the useful data from video footage for potential risk analysis at the urban scale. To address these challenges, this study proposes a new designed big data processing framework for potential traffic risk analysis and prediction. The proposed framework mainly consists of 1) a traffic video processing engine

and 2) potential traffic risk analysis and prediction. First, the traffic video processing engine aims to automate the process of extracting behavioral features from video and create one sequential process. Second, in the potential traffic risk analysis and prediction step, potential risk patterns are derived from the extracted vehicle’s and pedestrian’s behavioral features, and then it provides information on collision risks in advance through predicting their trajectories in real-time. In detail, the traffic video processing engine conducts partitioning video stream into video clips defined as a motioned-scene, detecting and segmenting traffic-related objects, generating their trajectories, and extracting their behavioral features in automatic as a one sequential process. The extracted behavioral features are stored in the form of a relation database for accessibility and ease of analysis of the data. Next, in the potential traffic risk analysis and prediction section, we first analyze vehicle-pedestrian interactions in snapshot level and scene level according to the characteristics of video. The snapshot level analysis handles the instantaneous interaction between vehicle and pedestrian by using data mining techniques to classify the severities of potential risk and derive rules about potential risks. On the other hand, in scene level analysis, we conduct to analyze behavioral feature changes by time and design a data cube model, called SafeyCube, based on a large amount of vehicle-pedestrian interactions in multiple spots and during the long-term period. In this analysis, we can conduct multi-dimensional analysis by OLAP operations with varying levels of abstractions. In addition, with visualizing the results, it makes easier for decision makers to obtain valuable information, such as potential risk patterns and risk-prone areas, enabling preemptive urban environmental improvements that do not depend on the accident history. Meanwhile, second part of the potential traffic risk analysis and prediction section is responsible for estimating a predictive potential risky area. In this part, we predict the trajectories of vehicle and pedestrian, and estimate the predictive risky boundaries based on deep learning and statistical inference. It can provide the collision warning in advance with driver and pedestrian through real-time manner. In addition, it proactively prevents traffic accidents rather than post-facto by measuring severity of potential risk in quantitative assessment. Furthermore, we believe that it can serve as part of safety services in C-ITS, next generation transportation system, by combining with communication technologies. The proposed framework in this dissertation enables decision-makers to gain a better understanding of how the vehicles and pedestrians behave near the crosswalk, and provide them with insights to improve the road environment safer.

Although the technological advancement of smart city infrastructure has significantly improved quality of our lives, road traffic accidents still pose a severe threat to human lives and have become a leading cause of premature deaths, making it a pressing current transportation concern. In particular, crosswalks are essential for pedestrians, but they are also a major threat. Nevertheless, we lack dense behavioral data to understand their risks faced by pedestrians crossing. In addition, the current traffic safety systems to protect the vulnerable road users (VRUs) rely on actual traffic accident history and collision statistics to determine the improvement of an urban environment post-facto. However, traffic accident events practically occur very rarely in usual traffic environments, so there is a limit to preventing traffic accidents by analyzing them. Therefore, it is necessary to derive analytical means for the safety assessment of pedestrians and the improvement of urban environment in proactive manner, not post-facto. Recently, with the rapid development of video processing and artificial intelligence technologies, much research has been conducted on vision-based traffic safety system using CCTVs. However, since most CCTVs record the video in oblique views, it is difficult to obtain object’s precise coordinates and behavioral features such as velocities and positions in automatics. Many studies have manually extracted them by human observations. Thus, it requires inefficiently costly and high time-consuming to do at the urban scale, so we seek to develop automated processes that generate useful data for potential risk analysis. To sum up, the current traffic safety systems have two issues: (1) a lack of analysis tools and prediction systems with the appropriate traffic accident data and various information to draw the solid and comprehensive conclusions for improving pedestrian safety