Lightweight graph transformers for clutter and target classification in automotive radar
Organizations: NXP Semiconductors, The Netherlands
Abstract
Radar is a core sensor modality for scene perception to achieve higher levels of autonomy in automotive driving. A common occurrence in automotive radars is clutter - detections of non-existent moving objects, that can adversely impact target detection and classification performance and subsequent driving actions. We propose lightweight tiny graph transformer network (TGTNet) models for classifying clutter from stationary and moving targets in the scene. Performance evaluation on the public RadarScenes show that our proposed TGTNet models achieve similar classification performance in precision and recall metrics in comparison to state of the art models, with one to two orders of magnitude lower model size and significantly faster inference.
Figures & tables
| Setup | LoRa | Model size | Inf. time (ms) | Moving object | Clutter | Stationary | Mean | ||||||||
| Precision | Recall | F1 | Precision | Recall | F1 | Precision | Recall | F1 | Precision | Recall | F1 | ||||
| GRT [ 13 ] | - | 20.07 MB | 25.47 | 90.76 | 91.47 | 91.12 | 95.42 | 95.32 | 95.37 | 99.88 | 99.85 | 99.86 | 95.35 | 95.55 | 95.45 |
| TGTNet (dim 32) | ✗ | 80.25 KB | 6.83 | 89.49 | 88.81 | 89.15 | 94.50 | 94.27 | 94.38 | 99.82 | 99.86 | 99.84 | 94.60 | 94.32 | 94.46 |
| TGTNet (dim 32) | ✓ | 40.29 KB | 7.10 | 88.33 | 89.29 | 88.81 | 94.66 | 93.85 | 94.25 | 99.83 | 99.83 | 99.83 | 94.28 | 94.33 | 94.30 |
| TGTNet (dim 16) | ✗ | 21.19 KB | 6.93 | 85.39 | 87.49 | 86.43 | 93.70 | 92.05 | 92.87 | 99.79 | 99.78 | 99.79 | 92.96 | 93.11 | 93.03 |
| TGTNet (dim 16) | ✓ | 11.21 KB | 7.39 | 85.33 | 87.15 | 86.23 | 93.21 | 91.95 | 92.58 | 99.79 | 99.77 | 99.78 | 92.78 | 92.96 | 92.86 |
| Setup | LoRa † | Model size | Inf. time (ms) | Moving object | Clutter | Stationary | Mean | ||||||||
| Precision | Recall | F1 | Precision | Recall | F1 | Precision | Recall | F1 | Precision | Recall | F1 | ||||
| S-TGTNet* (dim 16) | ✗ | 47.11 KB | 11.99 | 90.32 | 90.05 | 90.18 | 94.82 | 94.76 | 94.79 | 99.85 | 99.87 | 99.86 | 95.00 | 94.89 | 94.94 |
| SL-TGTNet* (dim 16) | ✗ | 14.88 KB | 10.56 | 88.85 | 85.63 | 87.21 | 92.95 | 93.20 | 93.07 | 99.75 | 99.87 | 99.81 | 93.85 | 92.90 | 93.36 |
| S-TGTNet* (dim 16) | ✓ | 27.78 KB | 12.80 | 90.11 | 89.04 | 89.57 | 94.27 | 94.91 | 94.59 | 99.85 | 99.86 | 99.85 | 94.74 | 94.60 | 94.67 |
| SL-TGTNet* (dim 16) | ✓ | 13.92 KB | 11.10 | 87.56 | 88.82 | 88.18 | 94.74 | 90.62 | 92.63 | 99.70 | 99.86 | 99.78 | 94.00 | 93.10 | 93.53 |
| S-TGTNet (dim 16) | ✓ | 28.03 KB | 12.01 | 90.92 | 89.39 | 90.15 | 94.80 | 92.67 | 93.72 | 99.71 | 99.89 | 99.80 | 95.14 | 93.98 | 94.56 |
| Setup | LoRa † | Model size | Inf. time (ms) | Moving object | Clutter | Stationary | Mean | ||||||||
| Precision | Recall | F1 | Precision | Recall | F1 | Precision | Recall | F1 | Precision | Recall | F1 | ||||
| [ 4 ] | - | 1.82 MB | 16.34 | 92.58 | 90.59 | 91.57 | 94.92 | 93.91 | 94.41 | 99.78 | 99.92 | 99.85 | 95.76 | 94.80 | 95.28 |
| [ 11 ] | - | 1.82 MB | 18.23 | 88.48 | 88.59 | 88.54 | 78.75 | 93.74 | 85.60 | 99.79 | 98.83 | 99.31 | 89.01 | 93.72 | 91.15 |
| GRU-TGTNet* | - | 27.15 KB | 30.76 | 87.90 | 88.49 | 88.19 | 94.25 | 93.36 | 93.80 | 99.82 | 99.84 | 99.83 | 93.99 | 93.90 | 93.94 |
| TGTNet-T (dim 16) | - | 11.52 KB | 10.26 | 88.39 | 89.83 | 89.10 | 94.72 | 93.78 | 94.25 | 99.85 | 99.83 | 99.84 | 94.32 | 94.48 | 94.40 |
| TGTNet-T (dim 32) | - | 40.91 KB | 9.94 | 90.21 | 91.26 | 90.73 | 95.66 | 94.68 | 95.17 | 99.86 | 99.87 | 99.87 | 95.24 | 95.27 | 95.26 |
| Setup | LoRa | Model size | Inf. time (ms) | Real | Ghost | Background | Mean | ||||||||
| Precision | Recall | F1 | Precision | Recall | F1 | Precision | Recall | F1 | Precision | Recall | F1 | ||||
| PointNet ++ | - | 3.74 MB | 11.26 | 94.67 | 84.20 | 89.13 | 43.72 | 66.79 | 52.85 | 99.81 | 99.93 | 99.87 | 79.40 | 83.64 | 80.62 |
| TGTNet (dim 16) | ✗ | 21.19 KB | 6.93 | 85.69 | 85.53 | 85.61 | 56.01 | 73.46 | 63.56 | 99.71 | 99.60 | 99.66 | 80.47 | 86.20 | 82.94 |
| TGTNet (dim 32) | ✗ | 80.25 KB | 6.83 | 89.82 | 81.09 | 85.23 | 54.00 | 70.39 | 61.11 | 99.61 | 99.77 | 99.69 | 81.14 | 83.75 | 82.01 |
| TGTNet (dim 16) | ✓ | 11.21 KB | 7.39 | 96.07 | 85.00 | 90.20 | 54.77 | 79.69 | 64.92 | 99.70 | 99.86 | 99.78 | 83.51 | 88.18 | 84.97 |
| S-TGTNet (dim 16) | ✓ | 28.03 KB | 12.01 | 94.87 | 80.26 | 86.96 | 52.66 | 68.14 | 59.41 | 99.55 | 99.88 | 99.71 | 82.36 | 82.76 | 82.03 |
Autonomous Driving Perception
Autonomous driving perception enables vehicles to detect and interpret road users, drivable space, and hazards using onboard and shared sensors.
See all 410 papers