Conclusion

Our simple snake algorithm performed nearly as well as the traditional methods in terms of accuracy. The snake algorithm had an accuracy rate of 0.817 compared to a rate of 0.840 for the traditional method, a loss in accuracy of only 0.023 This is very significant due to the much lower amount of work involved in using the snake. With the snake, one only has to mark the boundary of the desired feature on the first image in the sequence and the snake will do the rest, while for the traditional method, every image in the sequence has to be examined individually to pick out the feature. With large image sequences, the time savings due to using the snake algorithm is very significant.


We realized that careful tweaking of parameters is essential to the performance of the snake. Simply changing the number of points used (from 59 to 16), resulted in an increase in accuracy of 0.181 (from 0.524 to 0.705). With better choices of our parameters, it is highly likely that we can boost the performance of the snake to beyond the 0.840 of traditional methods.


With our current parameter settings, each control point tends to lock on and stay onto a particular point in the object. If the object is moving slowly, this is a boon, allowing for accurate tracking even under noisy conditions. However if the object changes shape too quickly, the point could become an edge for another object, and the snake would mistakenly follow the new object, affecting the accuracy of subsequent images.


We feel that it might be wise to provide the correct location of the feature every 10 cycles or so to ensure that any error does not get propagated for too long. While it would provide an increase in the work needed to use the snake, it would lead to great increases in accuracy. One could also then run the snake on the image sequence in reverse with the provided locations and interpolate the two results obtained for an even better match.


In conclusion, we are pleasantly surprised that our simple snake algorithm matches up pretty evenly with traditional methods in detecting a moving feature. With further work on our algorithm, we feel that the snake will prove to be much better than traditional methods in this field.