Traffic Light Triggering For Motorbikes Using Computer Vision Presentation

Vision Based Traffic Light Triggering for Motorbikes

Tommy Chheng tcchheng*at*ucsd.edu CSE 190: Projects In Computer Vision Instructor: Prof. Serge Belongie 15 March 07 1

Vision Based Traffic Light Triggering for Motorbikes ■

Goal: 

■

Motivation: 

■

To create a computer vision algorithm to detect incoming motorbike traffic for traffic light triggering. Current traffic light triggers do not work on motorbikes

General Idea: 

From video footage, track motion of objects and determine if the object's trajectory is intended incoming traffic.

2

Vision Based Traffic Light Triggering for Motorbikes ■

Motion Segmentation  

Background subtraction Defining the background image for subtraction: ➔

A still image from a set time -

➔

Previous Frame -

➔

Not robust Slight movement to the camera will produce noise. Difficult to automate updates. Difficult to pick up differences of slow moving objects.

Sliding Temporal Average -

Compute average frame from the previous n frames.

3

Vision Based Traffic Light Triggering for Motorbikes ■

Connected Component Labeling 



Label the foreground blobs for tracking. Accomplished by computing the distance from previous objects and attaching it to nearest one within a threshold. If not, create a new label.

4

Vision Based Traffic Light Triggering for Motorbikes ■

Lighting Conditions for Motion Segmentation 

Adversely affect motion segmentation



Use dual colorspace selection

During night time, RGB gives optimal results due to the headlights of the motorbikes and low noise. ➔ During day time, opponent color blue/yellow channel. Reduce effects of luminance. L*A*B* too nonlinear for low resolution images. ➔

Original

RGB Subtract

Opponent color B/Y Subtract

5

Vision Based Traffic Light Triggering for Motorbikes ■

Tracking 



Connected Component Labeling process labeled objects in each frame. Can save position information regarding each object.

Trajectories of objects in frame Objects labeled in frame 6

Vision Based Traffic Light Triggering for Motorbikes ■

Tracking Problems 





With simplified tracking determination, cross traffic can easily steal the intended tracking(or vice versa). Even if the tracking does not entirely work, partial tracking maybe enough to do the job. Use RANSAC to determine the partial trajectories.

Traffic mis-tracked

RANSAC to get partial trajectories

7

Vision Based Traffic Light Triggering for Motorbikes ■

Steps I did not have time for: 

Actual classification system ➔

➔



Ideas include feeding in annotated correct trajectories and testing by a difference on the test data.

Extensive training/testing set ➔



Now the traffic detection problem is reduced to a line fitting/classification problem.

Difficult to get obtain varied footage

Researched more on other methods at the various steps. ➔

Perhaps interest points moving in the same direction.

8