TABLE IV.
Comparison With Other Motion Estimation Works
| Algorithm | Sequence | Accuracy | Density | Implementation Platform | Frame Rate | |
|---|---|---|---|---|---|---|
| [28] | Motion Energy based | A bright spot on dark background, Translation (3, 0)1 | EE ≈ 0.2 pix.1 AE ≈ 1.2° 1 |
N/A | Analog 2μm ASICs + PC software | N/A (1 ms on ASICs) |
| A bright spot on dark background, Translation (4, 0)1 | EE ≈ 0.5 pix.1 AE ≈ 2.7° 1 |
|||||
| [35] | Motion Energy based | 9×9 white square on 128×128 dark background, Translation 1 pix/frame from 0° to 90° by 1° | EE = 0.12 pix.2 AE = N/A |
100% (16384 pix.) | FPGA Virtex-6 | 33 fps |
| [23] | Motion Energy based | 100×100 simple patches | EE = N/A AE = N/A |
100% (10,000 pix.) | Software on 14 PCs (933MHz) | 11 fps |
| [15] | Phase-based | 640×480 Yosemite | EE = N/A AE = 4.7° |
82.8% (254,360 pix.) | FPGA Virtex-4 (45 MHz) | 31.5 fps |
| [41] | Lucas-Kanade | 640×480 Yosemite | EE = 0.41 pix. AE = 8.7° |
0.21% (650 pix.) | FPGA Virtex-4 (62 MHz) + DSP DaVinci (800 MHz) | 160 fps |
| [42] | Multi-channel Gradient | 128×96 Yosemite | EE = N/A AE = 5.5° |
100% (12,288 pix.) | FPGA Virtex-2 | 16 fps |
| [43] | Tensor-based | 640×480 Yosemite | EE = N/A AE = 12.9° |
N/A | FPGA Virtex-2 Pro (100 MHz) | 64 fps |
| [44] | Block Matching (SAD-based) | 640×480 Yosemite | EE = N/A AE = N/A |
N/A | FPGA Virtex-2 | 30 fps |
| [45] | Lucas-Kanade | 800×600 Yosemite | EE = N/A AE = 3.2° |
61.6% (295,680 pix.) | FPGA Virtex-6 (94 MHz) | 196 fps |
| This work | Motion Energy based | 320×240 Yosemite, Translation (3,0) | EE = 0.29 pix. AE = 4.3° |
25% 3 (19,200 pix.) | FPGA Zynq-7020 (53 MHz) | 30 fps |
| 320×240 Yosemite, Translation (4,0) | EE = 0.39 pix. AE = 4.4° |
|||||
| 320×240 Yosemite, Translation 1 pix/frame from 0° to 90° by 45° | EE = 0.19 pix. AE = 5.8° |
|||||
| 320×240 Yosemite, Expansion 2% | EE = 0.35 pix. AE = 7.4° |
1
The original system is continuous-time and these values are renormalized assuming 1s contains 30 frames. Moreover, the EE and AE values are computed by ourselves by measuring the curves in Fig. 20 of [28] with an ideal assumption that the y-direction errors are always 0.
2
Only the error in horizontal (x-) direction is counted.
3
Even the pixels on boarders or with invalid confidence flag on the subsampled image are still processed to estimate its velocity on our system.