Author Affiliations
Department of Vehicle and Electrical Engineering, Army Engineering University Shijiazhuang Campus, Shijiazhuang, Hebei 050003, Chinashow less
Fig. 1. ComPoseNet model. (a) ComPoseNet workflow; (b) preprocessing and ComPoseNet architecture; (c) image processing
Fig. 2. Effect of loss function
Fig. 3. Geometric rule diagrams. (a) Selection rule schematic; (b) distance d; (c) angle θ; (d) angle α
Fig. 4. Effects of different loss functions. (a) Original image; (b) traditional method; (c) proposed algorithm; (d) comparison
Fig. 5. Translation errors of different algorithms
Fig. 6. Angle errors of different algorithms
Fig. 7. Effects of pose estimation of different algorithms
Fig. 8. Effects of different parameters on translation error
Fig. 9. Effects of different parameters on angle error
Fig. 10. Estimation effect diagrams of different parameter constraints. (a) Object; (b) d; (c) d+θ; (d) d+α; (e) d+θ+α; (f) comparison
Fig. 11. Target detection effects. (a) Telephone; (b) duck; (c) iron; (d) drill
Loss function | Number of training rounds | Translation error /m | Angle error /(°) | Pixel error | Accuracy /% |
---|
Traditional | 200 | 0.0459 | 8.310 | 5.982 | 86.391 | Proposed | 200 | 0.0425 | 7.731 | 5.576 | 91.262 |
|
Table 1. Errors and Accuracies of different loss functions
Algorithm | Number oftraining rounds | Translationerror /m | Angleerror /(°) |
---|
Traditional | 200 | 0.0459 | 8.310 | SSD | 200 | 0.0437 | 7.925 | YOLO | 200 | 0.0431 | 7.889 | Proposed | 200 | 0.0425 | 7.731 |
|
Table 2. Errors of different algorithms
Loss function | Number oftraining rounds | Translationerror /m | Angleerror /(°) |
---|
Traditional | 200 | 0.0459 | 8.310 | Proposed (d+θ) | 200 | 0.0442 | 8.125 | Proposed (d+α) | 200 | 0.0439 | 7.984 | Proposed(d+θ+α) | 200 | 0.0425 | 7.731 |
|
Table 3. Errors under different parameters