CS223B - Intro to Computer Vision Final Project

Test#4 - Lots of occlusion

As can be expected, a test with a lot of occlusion would yield problems for any vision system, especially if the ball changes directions in the occluded position.

Video:


Original Footage.

Final Results. Tracks the ball inconsistently because of repeated occlusions.

Comments:

This test case offers us some insight into potential improvements on the algorithm. The ball is successfully tracked for the first 25 frames or so. When the two players collide, the ball disappears behind one of the players and is lost. The program begins to track the players' heads for the next few frames. Although one might think that the player layer mask I had implemented would get rid of these effects, there is a lot of noise in these images and so not all of the player is masked off. The program continues to track one of the players until he approaches the ball to kick it at which the program locks on the ball again.

However, at this time another player comes in for a slide tackle and the ball is again occluded. The program "hangs out" on the field until the ball appears again and it tracks it until it goes behind another player. Observe this part carefully. The ball is so distant on the screen that the "+" completely covers the ball here. Unfortunately, when the ball reappears from behind the player the program has again lost the ball and is now in a confused region around the players. Notice how it clings on to a player's shirt, the line on the grass, a players leg, and a players arm as all potential "ball" manifestations. However, in the end it redeems itself by spotting the ball through the crowd and locking in on it just as a player approaches.