Project: #339 Vehicle Trajectory and Gap Estimation for Conflict Prediction Progress Report - Reporting Period Ending: Sept. 30, 2020 Principal Investigator: Alex Hauptmann Status: Active Start Date: July 1, 2020 End Date: June 30, 2021 Research Type: Applied Grant Type: Research Grant Program: FAST Act - Mobility National (2016 - 2022) Grant Cycle: 2020 Mobility21 UTC Progress Report (Last Updated: Oct. 5, 2020, 6:26 a.m.) % Project Completed to Date: 20 % Grant Award Expended: 0 % Match Expended & Document: 0 USDOT Requirements Accomplishments We are seeking to build a system to identify and measure vehicle trajectories, speed, inter-vehicle gaps, and vehicle pedestrian spacing using video streams from arbitrary traffic surveillance cameras. In the past period, we have followed our planned timeline and achieved solid progress. We have built an automatic camera parameter estimation model for roadway surveillance cameras based on simulation data and deep convolutional neural networks. With the calibrated parameters, we can calculate the world coordinates of traffic participants on the road plane. Built upon state-of-the-art object detection and multiple object tracking techniques, we have successfully developed a real-time detection-tracking system with 4 GPUs for full-HD traffic surveillance streams. Impacts The system that we have built is capable of accurately monitoring the speed, location, and trajectory of vehicles and pedestrians in the 3D reconstruction of the road plane. It runs in real-time for full HD traffic surveillance streams and is ready for deployment with its current set of functionalities. The derivatives of this project have demonstrated state-of-the- art performance on multiple challenges including the CVPR 2020 AI City Challenge and CVPR 2020 ActivityNet Activities in Extended Videos Challenge. We are actively sharing our latest results and solutions as open source with GPL licensing. We have also communicated our results to the Federal Highway Administration partners. We have proposed two monocular 3D object detection methods, one purely based on geometry, the other one leveraging the power of neural networks utilizing huge autonomous driving datasets. We have been actively advancing the state-of-the-art technology for activity detection in surveillance videos, including vehicles and pedestrians. We have developed a driving action detection system that uses no training data but obtains superior results compared to trained systems. Other Open-source software: Pyturbo: A pipeline system for efficient execution. 11k downloads since Jul. 2020. https://pypi.org/project/py-turbo/ AVI-R: A robust reader for AVI video files. 6.6k downloads since Jun. 2020. https://pypi.org/project/avi-r/ Outcomes New Partners None Issues None