PROJECT DESCRIPTION
The Vehicle Location by Thermal Image Features (VLTIF) system is an application designed to detect and count vehicles in traffic video sequences using thermal imaging cameras. Vehicle detection is an active and relevant research topic in computer vision and image processing. It is an essential ingredient in many traffic fields as much research has been focused on the optimization of traffic lights, enforcement of intersections, effectiveness of traffic reduction measures, and other traffic topics which rely on empirical evidence.
Current vehicle detection techniques rely almost exclusively on standard visible light spectrum cameras and have shown critical weaknesses in their detection accuracy. Whereas visible light cameras are sensitive to issues such as the time of day and reflection from surfaces, the use of thermal imaging should greatly improve the robustness and performance over standard visible light systems.
Our system will be stand-alone application which will take thermal imaging video as input and will output the results of the traffic detection system. The detection results will include a labeled video showing detected traffic as well as a formatted file denoting detection times and locations. This is useful for validation as well as for further processing by other systems. It will enable the user to open a video and label traffic lanes as lines. This will allow the system not only to count cars, but to organize them by traffic lanes.
Our system will include a graphical user interface which will enable the user to select relevant parameters as well as begin processing of video. The user will first select a video file to process. One major component is to allow the user to draw the traffic lanes onto the video display. By selecting the locations of traffic lanes, users will be allowed to count, segment, and track traffic flow in individual lanes. Once the lanes are established, the user will then begin processing.
The goal of our system is to reduce the complexity of traffic segmentation as much as possible. The user should have very little interaction with the system except to choose traffic lanes and begin processing. By reducing the complexity of our system into a "black box", we reduce the training time for novice users as well as reduce the likelihood for errors which occur from inappropriate configuration. This will require algorithms and techniques which are robust and light on variables. We have identified several techniques which will serve as a starting point to segmentation and classification. These techniques include mog, baysian classifier, and markov models. Our techniques will also require developing techniques for fusing fundamentally different imaging techniques.
VLTIF PRESENTATIONS AND POSTER
The first presentation covers the general problem or need that the project aimed to solve, with demonstations of proof-of-concept prototypes. The second presentation is an overview of the nearly finished project.
Project Video
PROJECT RESULTS
A picture -- or demo -- is worth a thousand words, and the following videos show the results of inputting traffic data collected in Reno by Team 11 into the VLTIF application. The final implemented VLTIF algorithm has been rigorously tested against nearly 60 minutes of ground truth data, with resulting accuracy rates above 92%.
EXTERNAL PROJECT RELATED RESOURCES
Existing Classifiers:
Below are two links to examples of existing vehicle classification and counting systems. Note that these appear to be excellent products: however, all examples shown are in daylight conditions with excellent visibility.
Excellent vehicle counting and classification during night time and inclement weather conditions are the end goals to separate VLTIF from other existing classification products.
Related Links:
- This project will draw on a rich set of existing image processing and computer vision techniques developed both theoretically and empirically over the last several decades.
(Book) Computer Vision: Algorithms and Applications