1.Introduction2.Previous Work3.Approach3.1.Motion Segmentation3.2.Tracking3.2.1Connected Component Labeling3.2.2Motion Tracking3.2.3RANSAC Partial Line Fitting4.Experimental Results5.ConclusionReferencesAbstractCurrent traffic light triggering is based on inductive loop sensors. Unfortunately, motorbikes (scooters, motorcycles, etc) have a difficult time triggering these sensors. In this paper, we propose an image processing algorithm to detect motorbikes at a traffic stop using a fixed camera. The algorithm tracks the trajectory of the objects in the footage by motion segmentation and connected component labeling. Classification can be created to categorize these objects as incoming traffic based on the object’s trajectory. To handle different lighting conditions in the motion segmentation, we take a dual approach by selecting RGB or Opponent colorspace. RANSAC is utilized to help trajectory creation. Experimental tests using real video footage exhibit robust results under varying conditions.1. IntroductionThis paper offers an algorithm to detect an incoming motorbike for traffic light triggering. The motivation is to solve a particularly dangerous situation for motorbikers. Motorbikes do not trigger the inductive loop sensors for the traffic lights and riders are left with little alternatives besides running the red light. This problem is most apparent at night time when less full size automobiles are present. Using a passive system such as a fixed camera connected to a computer system for processing has high potential as a solution. Cameras are non-obtrusive and can be robust in terms of maintenance and cost. To install or fix current inductive loop sensors, the road itself has to be closed. The road has to be torn to access to the sensors. On the other hand, a camera triggering system would be off the road and allow easy installation and servicing. A high mount placement would give the best results due to less occlusion by cross traffic. In some cases, many traffic intersections already have similar infrastructures in place with red light camera systems. There has been much work in the field of vehicle detection for traffic applications but very little work has been tested on motorbikes. The objective of this project is to demonstrate an algorithm to detect and track a motorbike for traffic light triggering. The methods used in this project are variations of conventional image processing techniques. We use image subtraction to obtain the motion segmentation of the footage. The background frame used is a sliding temporal average window. Accordingly, RGB colorspace is used for nighttime and the blue-yellow channel in the opponent colorspace is used for highly illuminated conditions such as daytime. A simple selection algorithm can select the colorspace based on a pixel intensity average. Binary thresholding and connected component labeling are applied to label the objects. RANSAC is used for partial line fittings to create robust trajectory lines. At this stage, we can classify objects as incoming traffic based on its trajectory. Ideally the techniques used in principle should work on automobiles. Some testing has been done on footage with automobiles but the main focus and parameter tuning has been set for motorbikes.Figure 1. Most motorcyclists are left stranded because current traffic light triggering sensors cannot detect their motorbikes. Vision Based Traffic Light Triggering for MotorbikesTommy ChhengDepartment of Computer Science and EngineeringUniversity of California, San [email protected]. Previous WorkMotion segmentation is a common image processing technique. Y. Liu et al. [1] presents a similar algorithm in segmenting automobiles from the background. This algorithm utilizes HSV colorspace mapping. We found using the blue yellow channel in the opponent colorspace proved more reliable than HSV for daytime illumination conditions. V. Bhuvaneshwar et al. [2] also utilizes a similar background subtraction for pedestrian detection.The other field of tracking involves tracking by feature points. The KLT tracker [4] is an example of this method.Our main goal was to utilize tracking solely for incoming traffic detection therefore a simple tracking based on motion segmentation and connected component labeling would be suitable.3. Approach3.1. Motion SegmentationThe motion segmentation algorithm applied is a variant of the common image subtraction technique. Image subtraction is often used to recover motion from a video footage. The background image is defined as all the static objects in the footage while the result of the image subtraction will give all the moving objects in the foreground. Defining the background image is the variation. Many projects define the background image to be the previous frame or a set empty frame. We define the background image to be a sliding temporal average window of the previous n frames. In our case, we set n to 15 frames. A sliding temporal average window proved to be most effective definition of the background image. Using just the previous frame gave low subtraction values when the motorbike was traveling at a slow rate. Additionally, using a set empty frame is possible but very susceptible to difficulties with camera vibrations or changing conditions. We used a dual algorithm for the image subtraction by selecting the colorspace. In late evening and nighttime, we used RGB subtraction. The headlights of motorbikes are defined cleanly against a dark background. In daytime, the lighting conditions are illuminated and motorbikes are difficult to distinguish from the background using RGB. We used only the blue/yellow channel in the opponent colorspace for daytime image subtraction. We discarded the luminance and red/green channel. RGB colorspace for image subtraction on daytime footage proved very ineffective at segmenting the motion of an incoming motorbike. Our first alternative was the (a)(b)(c)Figure 2. (a) Original video frame. (b) Image subtraction using RGB. (c) Image subtraction using the blue/yellow channel of Opponent colorspace.L*A*B* colorspace. The highly non-linear definitions of L*A*B* proved to be ineffective with low resolution imaging. We had also tested the HSV colorspace. We applied image subtraction on each component separately. The result was a noisy image. If we solely use the value (or