Author Affiliations
1 Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang, Liaoning 110016, China2 Institutes for Robotics and Intelligent Manufacturing, Chinese Academy of Sciences, Shenyang, Liaoning 110016, China3 University of Chinese Academy of Sciences, Beijing 100049, China4 Key Laboratory of Opto-Electronic Information Process, Chinese Academy of Sciences, Shenyang, Liaoning 110016, China5 Key Laboratory of Image Understanding and Computer Vision, Shenyang, Liaoning 110016, Chinashow less
Fig. 1. Schematic of ambiguity of original point pair feature. (a) Front view; (b) side view
Fig. 2. Visible constraint between viewpoints. (a) Schematic of visible constraint between two points; (b) 3D model; (c) 2.5D scene
Fig. 3. Viewpoint visibility constraint. (a) Location of point p; (b) variations in distribution and number of points satisfying visible constraint with point p with θ
Fig. 4. Flow chart of 3D object recognition based on enhanced point pair feature
Fig. 5. Dataset collected in practice. (a) Glass box and mouse model; (b) dataset R1; (c) dataset R2; (d) dataset R3
Fig. 6. Five models and two sample scenes of UWA dataset
Fig. 7. Recognition results of eight scenes on R1 dataset
Fig. 8. Time cost comparison between EPPF and original PPF based methods on R1 dataset
Fig. 9. Recognition results of four scenes S1-S4 on R2 dataset
Fig. 10. Time cost comparison on R2 dataset. (a) EPPF; (b) original PPF
Fig. 11. Recognition results of five scenes S1-S5 on R3 dataset
Fig. 12. Time cost comparison on R3 dataset. (a) EPPF; (b) original PPF
Model | Point number | Original PPF | Enhanced PPF |
---|
Number | Time cost /s | | Number | Time cost /s |
---|
Chef | 3351 | 11225850 | 85.10 | 5557422 | 34.70 | Chicken | 2643 | 6982806 | 44.00 | 3434958 | 18.90 | Para | 2507 | 6282542 | 43.20 | 3114468 | 17.50 | T-rex | 2337 | 5459232 | 35.90 | 2713202 | 14.80 | Glass Box | 1134 | 1284822 | 3.94 | 638622 | 1.87 | Mouse | 1358 | 1842806 | 6.16 | 916592 | 2.96 |
|
Table 1. Summary of basic information of six target models
Model | Correct number/total number | Failed scene number | Occlusion of the targets /% |
---|
Chef | 49/50 | 43 | 91.30 | Chicken | 45/48 | 6,26,32 | 89.70,86.50,89.50 | Parasaurolophus | 40/45 | 7,10,38,41,50 | 86.40,91.40,89.00,87.00,83.90 | T-rex | 41/45 | 4,10,34,48 | 84.00,80.20,83.80,77.30 | Average | 175/188 (93.1%) | - | - |
|
Table 2. Recognition results of proposed algorithm on whole UWA dataset
Algorithm | Recognition rate /% | Time cost for one object /s |
---|
Proposed | 97.6 | 10 | PPF in Ref. [8] | 97.0 | 85 |
|
Table 3. Comparison of two algorithms on the UWA dataset (occlusion of targets is less than 84%) in terms of recognition rate