Seite - 103 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

Evaluation. The evaluation aims at comparing our approach against the particle filter based approaches [5][18][9] and online learning based approach [20] in estimating the translation (in x, y, and z axis) and rotation (roll, pitch and yaw) parameters. We compute the root mean square (RMS) errors (translation, rotation) and average time per frame. Table I shows our approach outperforms [5] and [18] over all sequences. Unlike [5] our approach only uses depth data for 3D tracking. It also performs better than [9] on average of about 0.31 mm and 1.16 deg in estimating the translation and rotation parameters respectively. Our approach requires much less computational time (1.7 ms per frame) when compared with [9] (131 ms). Though our approach performs on par with [20] in terms of run-time, it performs better in estimating the translation (by 0.31 mm) and rotation parameters (by 0.15 deg) on average. TABLE I. COMPARISION OF OUR APPROACH WITH THE STATE OF ART AGAINST THE RMS ERRORS IN TRANSLATION (IN MM), ROTATION (DEGREES) AND THE RUNTIME (MS) PCL [18]1 Choi [5]2 Krull [9]3 Tan [20]4 Ours5 Transl. (x) 13.38 0.93 0.51 1.23 0.63 Transl. (y) 31.45 1.94 1.27 0.74 1.19 Transl. (z) 26.09 1.09 0.62 0.24 0.48 Roll 59.37 3.83 2.19 0.50 0.19 Pitch 19.58 1.41 1.44 0.28 0.28 Yaw 75.03 3.26 1.90 0.46 0.27 Time 2205 134 135 1.5 1.7 Transl. (x) 2.53 0.96 0.52 1.10 0.39 Transl. (y) 2.20 1.44 0.74 0.94 0.37 Transl. (z) 1.91 1.17 0.63 0.18 0.37 Roll 85.81 1.32 1.28 0.35 0.12 Pitch 42.12 0.75 1.08 0.24 0.17 Yaw 46.37 1.39 1.20 0.37 0.15 Time 1637 117 129 1.5 1.69 Transl. (x) 1.46 0.83 0.69 0.73 0.42 Transl. (y) 2.25 1.37 0.81 0.56 0.51 Transl. (z) 0.92 1.20 0.81 0.24 0.64 Roll 5.15 1.78 2.10 0.31 0.22 Pitch 2.13 1.09 1.38 0.25 0.29 Yaw 2.98 1.13 1.27 0.34 0.30 Time 2762 111 116 1.5 1.7 Transl. (x) 43.99 1.84 0.83 1.54 0.30 Transl. (y) 42.51 2.23 1.67 1.90 0.49 Transl. (z) 55.89 1.36 0.79 0.34 0.31 Roll 7.62 6.41 1.11 0.42 0.21 Pitch 1.87 0.76 0.55 0.22 0.27 Yaw 8.31 6.32 1.04 0.68 0.23 Time 4539 166 143 1.5 1.71 Transl. 18.72 1.36 0.82 0.81 0.50 Rot. 29.70 2.45 1.38 0.37 0.22 Time 2786 132 131 1.5 1.7 1,2 Intel Core2 Quad CPU Q9300, 8G RAM with Nvidia GTX 590 GPU; 3 Intel(R) Core(TM) i7 CPU with a Nvidia GTX 550 TI GPU; 4 Intel(R) Core(TM) i7 CPU; 5 Intel(R) Core(TM) i5 CPU 6. Conclusion We have presented a framework for combining object tracking and object localization to provide robust tracking performance in a challenging scenario. A quantitative analysis of the evaluation on popular test data set is also presented. The evaluation shows that our approach performs better than the state of art in terms of estimating the translation and rotation parameters. The approach is 103