United States Department of Transportation
i
Assessment of Contextual Complexity and Risk Using Unsupervised Clustering Approaches with Dynamic Traffic Condition Data Obtained from Autonomous Vehicles
-
2024-05-01
-
[PDF-1.41 MB]
Details:
-
Alternative Title:A Statistical and Machine Learning Approach to Assess Contextual Complexity of the Driving Environment Using Autonomous Vehicle Data [Title from Cover]
-
Creators:
-
Corporate Creators:
-
Corporate Contributors:
-
Subject/TRT Terms:
-
Resource Type:
-
Geographical Coverage:
-
Edition:Final Report (August 2021 - May 2024)
-
Corporate Publisher:
-
Abstract:Traditional road safety assessment methodologies rely heavily on AADT (Annual Average Daily Traffic) data estimates to account for variability in traffic operations. Unfortunately, this approach does not consider the driving environment's fast-changing dynamics, which can influence contextual complexity and risk. This research report presents a method to measure and quantify the contextual complexity of the roadway environment using diverse open-source LiDAR (Light Detection and Ranging) sensor data collected by Waymo autonomous vehicles under dynamic traffic conditions. The proposed Contextual Complexity Factor (CCF) model estimates the driving scene's complexity using the density and proximity of the objects around the vehicle. Besides, an unsupervised machine learning technique using clustering algorithms was used to measure and classify the driving environment's dynamic characteristics (e.g., vehicles, pedestrians, bicycles) into appropriate risk categories to develop a dynamic complexity model. Variables, including velocity, object density of lidar, and object proximity, were selected for k-means and hierarchical clustering analysis. Three clusters were ultimately chosen that categorize the scene into high, medium, and low categories of complexity. Adopting the results from the clustering analysis, the research team further built the complexity ranges for the attributes (i.e., velocity, object density, and object proximity). Both statistical and machine learning models were proficient in predicting the dynamic complexity with justifiable truthfulness. Identifying and predicting high-risk environments in real-time can significantly benefit safety research, driver education, auto-insurance risk assessment, autonomous vehicle route planning, and many more.
-
Format:
-
Funding:
-
Collection(s):
-
Main Document Checksum:
-
Download URL:
-
File Type:
