Comparison of OpenCV Tracking Algorithms for a Post-Stroke Rehabilitation Exergame

• Júlia Tannús UFU

Resumo

In this paper, a comparison was made between three OpenCV algorithms: 1) HSV color threshold, 2) Discriminative Correlation Filter with Channel and Spatial Reliability (CSRT) and 3) Mean shift. The objective is to use one of them in an exergame for the rehabilitation of gross movements of the upper limb post-stroke. In the game, the patient would hold an object, the position of which would be tracked by one of these algorithms through a common RGB camera and used as an input to the exergame. Tests were performed for success, accuracy and time per frame. This exact comparison has not been found previously in the literature. The success comparison represented the percentage of frames that had a tracked and ground truth region of interest overlap above 50%. The CSRT algorithm had a median success of 1.00, or perfect, followed by the color algorithm, which scored 0.95 and the Mean shift, with 0.48. The precision comparison measured the intersection over union value in relation to a manually adjusted ground truth. The color algorithm was the most precise of all, with a median of 0.88, followed by the CSRT algorithm, with 0.74 and the Mean shift, with 0.41. The time per frame comparison showed that the color and Mean shift algorithms reached 140.8 fps and 144.9 fps, respectively, while the CSRT obtained 12.6 fps, a value similar to what the authors claimed. Thus, the color algorithm was the most reliable, had the lowest processing cost and was the chosen one for the application.

Palavras-chave: stroke rehabilitation, color tracking, CSRT, Mean shift, OpenCV