Ball Detection and Results

Although the process of generating the cascade of classifiers is relatively time-consuming, it produces quite promising results. The first three stages of a learned cascade are shown in Fig. 6. The cascade was tested systematically using four categories of input data (see Fig. 7):

1. RoboCup football arena
2. Uniform background
3. Rescue robotics arena
4. Complex scenes

The idea behind this categorization was to be able to determine exactly where the strengths and weaknesses of the technique lie. The algorithm is to be applied to football-playing robots participating in RoboCup to detect footballs, the innovation being the independence of color and surface pattern of the ball. The first test to be passed was therefore detecting the three different balls in the RoboCup arena. The algorithm not only works well in the arena, but with any images with a uniform background behind the footballs. This is shown by the next test set. To test our method to its limits we tested it both with images taken in the RoboCup Rescue arena, as well as other more complex images, with all sorts of noise and distracting objects. For each image category we ran the test with 25 pictures, including 15 of each ball, making a total of 100 images and a total of 180 balls (60 each). The detection speed averaged at 300msec per image. A sample of the test images in each category is shown in Fig. 7; the results of the experiments run on two different cascades are given in Tables I and II. The tables reveal how many red, white or yellow balls were correctly classified or not detected, as well as the number of false positives in each image category. The first cascade was learned with a total of 1000 images, where the second used a mere 570 input images.

Figure: Top left: RoboCup football arena. Top right: Uniform background. Bottom left: RoboCup Rescue arena. Bottom right: Complex images.

Table I: Simple Training Data

		Correct		Not Detected		False Pos.
	red	13		2
RoboCup	wht	5	26/45	10	19/45	4
	yel	8		7
	red	13		2
Uniform	wht	13	39/45	2	6/45	0
	yel	13		2
	red	2		13
Rescue	wht	2	5/45	13	40/45	10
	yel	1		14
	red	0		15
Complex	wht	0	0/45	15	45/45	9
	yel	0		15
	red	28/60		32/60
Total	wht	20/60		40/60		23
	yel	22/60		38/60

Table: Complex Training Data

		Correct		Not Detected		False Pos.
	red	14		1
RoboCup	wht	11	37/45	4	8/45	5
	yel	12		3
	red	15		0
Uniform	wht	14	42/45	1	3/45	4
	yel	13		2
	red	5		10
Rescue	wht	5	20/45	10	25/45	21
	yel	10		5
	red	5		10
Complex	wht	2	9/45	13	36/45	24
	yel	2		13
	red	39/60		21/60
Total	wht	32/60		28/60		54
	yel	37/60		23/60

Actual detected images are shown in Fig. 8 for all four categories, for both cascades, where the balls detected by the simple cascade only are marked by blue boxes, those detected by the complex cascade by green boxes and those detected by both by red boxes.

Figure: Sample input images of each of the four categories (From top to bottom: RoboCup, uniform background, RoboCup Rescue and complex scenes) and the corresponding results of the two classifiers.

The two classifiers used in the experiments differ only in the data used for training. For the first classifier (Table I) relatively simple images were used for training, where the images contained no background noise. The idea behind this approach was to make sure that the classifier would only use information about the ball itself and that complex data would only be confusing and would eliminate all useful information about the ball shape. For the second classifier (Table II) a wider range of training data was used, including images with different lighting, in different surroundings (similar to the complex images used for testing), etc. The difference in performance is quite clear from the detected images (Fig. 8).

It can also be observed from the results that we will always face the difficulty of false positives when differentiating between footballs and any other round objects. Examples of such false positives are shown in Fig. 9.

Figure 9: Left: Input images including round objects. Right: False detections in filtered images.