Background
SwRI has previously developed a traffic camera vehicle analytics software called Active-Vision. While pursuing various proposals to deploy Active-Vision, it was found that many of these proposals either require or strongly desire pedestrian detection as a feature in these proposals along with vehicle analytics. To pursue these proposals, our research team researched the technologies and approaches necessary to detect pedestrians using traffic cameras.
Approach
While person detection has been well explored in previous areas such as autonomous vehicles and robotics, these areas tend to have a much closer, low to the ground, and high-resolution view of the people. Traffic cameras tend to be far away from what they are viewing, mounted 40+ feet up, and are typically low resolution. This causes the persons in the image to represent a small number of pixels making traditional machine learning approaches infeasible for detecting pedestrians in traffic cameras. To solve this, we explored using background subtraction techniques which do not have the image resolution issues that machine learning algorithm do. While background subtraction was able to find the pedestrians, it also resulted in a lot of extra noise being produced which caused false positives. Much of the research was around how to characterize the type of noise seen in traffic camera perspectives and to create techniques for suppressing this noise.
Accomplishments
The result of the project was an algorithm that had greater than 80+ percent accuracy at predicting pedestrians on highway cameras. This was tested on over 10 different cameras across various times of day and weather conditions.