Author Affiliations
1School of Electronics and Information Engineering, Suzhou University of Science and Technology, Suzhou, Jiangsu 215009, China2Tsinghua University Suzhou Automotive Research Institute, Suzhou, Jiangsu 215134, Chinashow less
Fig. 1. Algorithm frame diagram
Fig. 2. Mask prediction module
Fig. 3. Prediction module
Fig. 4. Point cloud projection
Fig. 5. Calibration board effect
Fig. 6. Mask projection
Fig. 7. Input sequence RGB and mask data into the model. (a) Original sequence RGB data; (b) Corresponding sequence mask data
Fig. 8. Loss function curve. (a) Four definitions of loss; (b) Total algorithm loss
Fig. 9. PR curve comparison
Fig. 10. The accuracy of the three algorithms for different distance, occlusion, luminosity and ambiguity
Fig. 11. Performance of multiple target tracking algorithms on MOT
Fig. 12. Effect of Apollo dataset test
Fig. 13. Effect of KITTI dataset test
Fig. 14. Effect of BDD100K dataset test
Fig. 15. Effect of point cloud projection
Fig. 16. Experimental platform
Fig. 17. Effect of actual road experiment
Method | Backbone | ms | rc | Epochs | AP | AP50 | AP75 | APS | APM | APL | APbb | fps | Mask R-CNN [19] | R-50-FPN | | | 12 | 34.6 | 56.5 | 36.6 | 15.3 | 36.3 | 49.7 | 38.0 | 8.6 | Mask R-CNN | R-101-FPN | | | 12 | 36.2 | 58.6 | 38.5 | 16.4 | 38.4 | 52.0 | 40.1 | 8.1 | Mask R-CNN | R-101-FPN | √ | | 36 | 38.1 | 60.9 | 40.7 | 18.4 | 40.2 | 53.4 | 42.6 | 8.7 | YOLACT-700 [21] | R-101-FPN | √ | √ | 48 | 31.2 | 50.6 | 32.8 | 12.1 | 33.3 | 47.1 | - | 23.6 | Ours | R-50-FPN | | | 12 | 33.6 | 54.5 | 35.4 | 15.1 | 35.9 | 47.3 | 38.2 | 16.7 | Ours | R-101-FPN | √ | | 36 | 37.7 | 59.1 | 40.3 | 17.9 | 40.4 | 53.0 | 42.5 | 13.7 | Ours-600 | R-101-FPN | √ | | 36 | 35.2 | 55.9 | 37.3 | 12.4 | 37.3 | 54.9 | 40.2 | 21.7 |
|
Table 1. This algorithm is compared with other algorithms
| Method | sMOTSA | MOTSA | MOTSP | KITTI mots dataset-cars | Mask R-CNN | 74.9 | 85.8 | 85.1 | MaskTrackR-CNN [25] | 75.5 | 86.1 | 86.5 | Track R-CNN [26] | 76.2 | 86.8 | 87.2 | Ours | 77.6 | 87.8 | 86.3 | KITTI mots dataset-pedestrains | Mask R-CNN | 44.6 | 63.8 | 74.1 | MaskTrack R-CNN | 45.9 | 64.6 | 77.9 | Track R-CNN | 46.8 | 65.1 | 75.7 | Ours | 45.3 | 65.6 | 77.0 |
|
Table 2. Performance index