Ensure That Video-based Automated Pedestrian and Cyclist Counting Systems Can Handle Varying Scenarios and Multi-Object Detection.
Louisiana Study Identified Contributing Factors for Low Accuracy from an Automated Video-Based Counting System.
Made Public Date
11/29/2021
TwitterLinkedInFacebook
Identifier
2021-L01070

ITS Support for Pedestrians and Bicyclists Count: Developing a Statewide Multimodal Count Program

Background

Manual counting techniques for gathering pedestrian and cyclist volume data are inefficient and impractical for large scale counting programs. As a result, there is great interest in accurate automated data collection techniques. Researchers in Louisiana investigated an automated pedestrian and cyclist counting system that analyzes archived video footage. The Histogram of Oriented Gradient technique, an open source method with high speed processing capability, was used to process video images to count pedestrians and cyclists. The accuracy of the automated method was evaluated against manual counts of pedestrians and cyclists, using 2017 video from five locations in Baton Rouge, Louisiana with varying levels of pedestrian and cyclist density.  The results showed that accuracy rates of the automated detection method ranged between 29 and 91 percent for pedestrians and between 0 and 60 percent for cyclists. Lower accuracy rates were observed in the presence of higher density of pedestrians.

Lessons Learned

  • Enhance viewpoint of cameras. Viewpoint angle was a major factor affecting the accuracy of the automated method. A consistent angle of camera mounting should be maintained for better accuracy.
  • Detection techniques should account for varying circumstances. Different light intensities, video time periods, motion patterns for tracked objects, and complex backgrounds should be considered in the development of accurate multi-object detection techniques.
  • Use a dataset with varied viewpoints to train the object detection model. Models should provide better accuracy rates if trained with a dataset that has various viewpoints including true positive cases with pedestrians and cyclists as well as false positives containing background trees, buildings, and other objects.
  • Add pedestrian tracking and cyclist tracking to the algorithm for counting. Lighting conditions can cause inaccurate detection. Detection accuracy can be improved by capturing and storing the location of the object over successive video frames.
  • High speed algorithms should be chosen for automated online counting. An accurate multi-object tracking algorithm is required to minimize false positives in the count process, but this operation can be time-consuming. High speed algorithms are required to support an online system for automated pedestrian and cyclist counting.
Goal Areas