File size: 27,067 Bytes
ae419ed | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 | Dưới đây là **kế hoạch hoàn chỉnh phiên bản mạnh hơn** cho paper phát hiện té ngã, có dataset link, mô tả dataset, method lấy từ các paper khác, proposed method, thí nghiệm, bảng kết quả và lộ trình chạy.
# 1. Hướng paper chốt
## Tên paper đề xuất
**DynaFall: Robust Skeleton-Based Fall Detection via Keypoint Dynamics and Graph-Based Benchmarking**
Tên này tốt hơn tên cũ vì nó thể hiện 2 điểm:
1. Có **model riêng**: DynaFall.
2. Có **benchmark với các method mạnh từ paper khác**: ST-GCN, 2s-AGCN, CTR-GCN, PoseC3D, TCNTE-style.
Ý tưởng chính:
```text
RGB Video
→ YOLO Pose
→ 17-keypoint skeleton sequence
→ Joint / Bone / Dynamics features
→ Compare with strong skeleton methods
→ Proposed DynaFall-GCN
→ Fall / Non-fall
```
Paper sẽ không còn đơn giản là “dùng CNN/LSTM/Transformer để phân loại té ngã”, mà thành:
> Một framework phát hiện té ngã dựa trên skeleton, chạy lại các method skeleton action recognition mạnh trên cùng dataset, sau đó đề xuất model chuyên biệt cho fall detection bằng keypoint dynamics và confidence-aware dropout.
---
# 2. Dataset dùng trong paper
Tôi khuyên dùng **2 dataset chính + 1 dataset backup/extension**.
## Dataset chính 1: UR Fall Detection Dataset — URFD
**Link dataset:** UR Fall Detection Dataset. ([Fenix][1])
**Mô tả:** URFD gồm **70 sequences**, trong đó có **30 fall sequences** và **40 activities of daily living — ADL sequences**. Fall events được ghi bằng **2 Microsoft Kinect cameras** kèm dữ liệu accelerometer; ADL events được ghi bằng camera 0 và accelerometer. ([Fenix][1])
**Vì sao nên dùng:**
* Nhỏ, dễ chạy nhanh.
* Rất phổ biến trong fall detection.
* Phù hợp để debug pipeline trước.
* Có RGB video để trích xuất YOLO pose.
**Vai trò trong paper:**
```text
Dataset kiểm tra nhanh + benchmark chính.
```
---
## Dataset chính 2: Multiple Cameras Fall Dataset — MCFD
**Link dataset:** Multiple Cameras Fall Dataset. ([Iro Montéal][2])
**Mô tả:** MCFD có **24 scenarios** được ghi bằng **8 IP video cameras**. **22 scenarios đầu** chứa fall và các sự kiện gây nhầm lẫn; **2 scenarios cuối** chỉ chứa các sự kiện gây nhầm lẫn. ([Iro Montéal][2])
**Vì sao nên dùng:**
* Có nhiều góc camera.
* Phù hợp để kiểm tra robustness.
* Paper nhìn mạnh hơn URFD-only.
* Có các tình huống dễ gây nhầm: ngồi, cúi, nằm, chuyển động gần giống fall.
**Vai trò trong paper:**
```text
Dataset kiểm tra multi-view + robustness.
```
---
## Dataset backup/extension: UP-Fall Detection Dataset
**Link dataset:** UP-Fall / Challenge UP data page. ([Google Sites][3])
**Mô tả:** Trang Challenge UP mô tả UP-Fall là dataset lớn cho fall detection, gồm **11 activities**, **3 trials per activity**, thực hiện bởi **12 subjects**, bao gồm 6 hoạt động hằng ngày và 5 kiểu té ngã, thu bằng wearable sensors, ambient sensors và vision devices. ([Google Sites][3]) Paper gốc về UP-Fall cũng nhấn mạnh đây là dataset multimodal cho fall detection và human activity recognition. ([MDPI][4])
**Khi nào dùng:**
* Nếu MCFD tải khó hoặc xử lý annotation khó.
* Nếu muốn paper mạnh hơn và có thêm dataset thứ 3.
* Nếu muốn hướng “multimodal dataset nhưng chỉ dùng RGB-derived pose”.
**Khuyến nghị thực tế:**
Paper 8 trang cho hội nghị Q4 thì dùng **URFD + MCFD** là đủ. UP-Fall để backup.
---
# 3. Dataset protocol chốt
Trong paper ghi rõ:
```text
Although some datasets provide depth, accelerometer, or multimodal signals, this study only uses RGB videos to extract human skeleton keypoints. This ensures a consistent vision-based evaluation protocol across all datasets.
```
Tức là:
| Dataset | Dữ liệu gốc có gì | Mình dùng gì |
| ---------------- | ------------------------- | -------------------------- |
| URFD | RGB, depth, accelerometer | Chỉ RGB → YOLO Pose |
| MCFD | Multi-camera RGB video | RGB → YOLO Pose |
| UP-Fall nếu dùng | wearable, ambient, vision | Chỉ RGB/vision → YOLO Pose |
Điểm này rất quan trọng để reviewer thấy so sánh công bằng.
---
# 4. Proposed method
## Tên model
**DynaFall-GCN**
## Ý tưởng
Model của mình không chỉ dùng skeleton thô, mà có 3 stream:
```text
Stream 1: Joint stream
- x, y, confidence của 17 keypoints
Stream 2: Bone stream
- vector xương giữa các keypoints
Stream 3: Dynamics stream
- velocity
- acceleration
- torso angle
- hip drop
- body aspect ratio
- center-of-body motion
```
Sau đó dùng:
```text
Graph-temporal encoder
+ confidence-aware keypoint dropout
+ classification head
```
---
# 5. Pipeline tổng thể
```text
Input RGB Video
↓
Frame Sampling
T = 32 frames
↓
YOLOv8/YOLO11 Pose Extraction
17 COCO keypoints per frame
↓
Pose Normalization
bbox-centered normalization
↓
Feature Construction
joint + bone + dynamics
↓
Comparative Methods
ST-GCN / 2s-AGCN / CTR-GCN / PoseC3D / TCNTE-style
↓
Proposed DynaFall-GCN
joint stream + bone stream + dynamics stream
↓
Robustness Evaluation
missing keypoints / lower-body occlusion / low-confidence mask
↓
Fall / Non-fall Prediction
```
---
# 6. Các method từ paper khác để chạy lại
Đây là phần giúp paper “hay hơn” và không bị đơn giản.
## 6.1 LSTM baseline
Dùng để kiểm tra sequence baseline.
Input:
```text
32 frames × 17 keypoints × 3
→ flatten
→ 32 × 51
```
Vai trò:
```text
Basic temporal baseline.
```
---
## 6.2 ST-GCN
**Paper:** Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition, AAAI 2018.
ST-GCN mô hình hóa skeleton như một graph không-thời gian, trong đó keypoints là nodes, bones là edges, và mô hình học cả quan hệ không gian giữa khớp và quan hệ thời gian giữa frames. ([AAAI Publications][5])
Áp dụng vào paper:
```text
Input: N × C × T × V × M
C = 3
T = 32
V = 17
M = 1
Output = Fall / Non-fall
```
Vai trò:
```text
Classic graph-based skeleton baseline.
```
---
## 6.3 2s-AGCN
**Paper:** Two-Stream Adaptive Graph Convolutional Networks for Skeleton-Based Action Recognition, CVPR 2019.
2s-AGCN mạnh hơn ST-GCN vì graph topology có thể được học thích nghi, không cố định hoàn toàn. Method này cũng dùng hai stream: **joint stream** và **bone stream**, trong đó bone stream chứa thông tin độ dài và hướng xương. ([arXiv][6])
Áp dụng:
```text
Joint stream: keypoint coordinates
Bone stream: bone vectors
Final prediction: score fusion
```
Vai trò:
```text
Adaptive graph + bone information baseline.
```
---
## 6.4 CTR-GCN
**Paper:** Channel-wise Topology Refinement Graph Convolution for Skeleton-Based Action Recognition, ICCV 2021.
CTR-GCN học graph topology theo từng channel, giúp mô hình học các quan hệ khớp khác nhau cho từng loại feature. Đây là một baseline GCN mạnh cho skeleton action recognition. ([arXiv][7])
Áp dụng:
```text
Input: joint + bone skeleton sequence
Graph: COCO 17-keypoint topology
Output: Fall / Non-fall
```
Vai trò:
```text
Strong graph-based baseline.
```
---
## 6.5 PoseC3D / PoseConv3D
**Paper:** Revisiting Skeleton-based Action Recognition, CVPR 2022.
PoseC3D/PoseConv3D không dùng graph sequence trực tiếp, mà chuyển pose thành **3D heatmap volume**. Paper gốc cho rằng cách này hiệu quả hơn trong học spatiotemporal features, robust hơn với pose estimation noise và generalize tốt hơn trong cross-dataset settings. ([CVF Open Access][8])
Áp dụng:
```text
YOLO keypoints
→ 3D pose heatmap volume
→ 3D CNN
→ Fall / Non-fall
```
Vai trò:
```text
Heatmap-based skeleton baseline.
```
---
## 6.6 TCNTE-style
**Paper:** Real-time skeleton-based fall detection algorithm based on Temporal Convolutional Networks and Transformer Encoder, 2025.
TCNTE kết hợp **Temporal Convolutional Network** với **Transformer Encoder** cho fall detection thời gian thực, dùng YOLOv8 pose extraction, BoT-SORT tracking, và weighted focal loss để xử lý mất cân bằng lớp. ([ScienceDirect][9])
Áp dụng:
```text
TCN + Transformer Encoder
Input: skeleton sequence
Loss: weighted focal loss
Output: Fall / Non-fall
```
Vai trò:
```text
Recent fall-specific temporal baseline.
```
---
# 7. Bảng method chính trong paper
Nên đưa bảng này vào phần **Experimental Setup**.
| Method | Loại | Nguồn gốc | Input | Vai trò |
| ---------------- | ----------------- | ------------------- | ----------------------- | --------------------------- |
| LSTM | Sequence model | baseline phổ biến | joint sequence | baseline cơ bản |
| ST-GCN | Graph-based | AAAI 2018 | joint graph | skeleton baseline kinh điển |
| 2s-AGCN | Adaptive GCN | CVPR 2019 | joint + bone | baseline mạnh hơn ST-GCN |
| CTR-GCN | Refined GCN | ICCV 2021 | joint + bone | strong GCN baseline |
| PoseC3D | Heatmap 3D CNN | CVPR 2022 | pose heatmap | khác hướng GCN |
| TCNTE-style | TCN + Transformer | fall detection 2025 | skeleton sequence | fall-specific baseline |
| **DynaFall-GCN** | Proposed | ours | joint + bone + dynamics | model chính |
---
# 8. Điểm mới của DynaFall-GCN so với các method kia
| Thành phần | ST-GCN | 2s-AGCN | CTR-GCN | PoseC3D | TCNTE | DynaFall-GCN |
| --------------------------------- | -------------- | -------------- | -------------- | --------------- | ------------------------- | ------------ |
| Joint stream | Có | Có | Có | Heatmap | Có | Có |
| Bone stream | Không | Có | Có | Không trực tiếp | Không rõ | Có |
| Velocity/Acceleration | Không | Không | Không | Không trực tiếp | Có thể có motion temporal | Có rõ |
| Torso angle | Không | Không | Không | Không | Không | Có |
| Hip drop | Không | Không | Không | Không | Không | Có |
| Body aspect ratio | Không | Không | Không | Không | Không | Có |
| Confidence-aware dropout | Không | Không | Không | Không | Không rõ | Có |
| Robustness test missing keypoints | Không bắt buộc | Không bắt buộc | Không bắt buộc | Có thể robust | Không bắt buộc | Có |
Thông điệp chính:
> Các method skeleton action recognition mạnh học action nói chung, còn DynaFall-GCN đưa thêm fall-specific dynamics và robustness training để phù hợp hơn với bài toán té ngã.
---
# 9. Chi tiết xử lý dữ liệu
## Step 1: Download dataset
Tải:
1. **URFD** từ trang chính thức. ([Fenix][1])
2. **MCFD** từ trang dataset của Université de Montréal. ([Iro Montéal][2])
3. **UP-Fall** chỉ dùng nếu cần backup hoặc mở rộng. ([Google Sites][3])
---
## Step 2: Chuẩn hóa thư mục
```text
fall_project/
├── data/
│ ├── raw/
│ │ ├── URFD/
│ │ ├── MCFD/
│ │ └── UPFall_optional/
│ ├── poses/
│ ├── clips/
│ └── splits/
├── src/
├── configs/
├── results/
└── paper/
```
---
## Step 3: Trích xuất pose
Dùng:
```text
YOLOv8n-pose hoặc YOLO11n-pose
```
Output mỗi frame:
```text
17 keypoints × 3
x, y, confidence
```
Nếu có nhiều người trong frame:
```text
Chọn người có bbox lớn nhất.
```
Lý do:
```text
Fall datasets thường chỉ có một subject chính.
```
---
## Step 4: Normalize keypoints
Với mỗi frame:
```text
x_norm = (x - bbox_center_x) / bbox_width
y_norm = (y - bbox_center_y) / bbox_height
confidence giữ nguyên
```
Output:
```text
T × 17 × 3
```
Trong đó:
```text
T = 32 frames
```
---
## Step 5: Tạo clip
Dùng:
```text
clip length = 32 frames
stride = 16
```
Nếu video ngắn:
```text
uniform sampling hoặc padding frame cuối
```
Quan trọng:
```text
Chia train/val/test theo video trước, rồi mới tạo clip.
Không random clip trực tiếp vì sẽ leak dữ liệu.
```
---
# 10. Feature engineering cho proposed model
## 10.1 Joint features
```text
x, y, confidence
```
Shape:
```text
32 × 17 × 3
```
---
## 10.2 Bone features
Dựa trên COCO skeleton edges:
```text
bone_vector = child_joint - parent_joint
```
Các cạnh chính:
```text
nose-eye-ear
shoulder-elbow-wrist
shoulder-hip
hip-knee-ankle
left-right shoulder
left-right hip
```
Shape:
```text
32 × 17 × 3
```
---
## 10.3 Dynamics features
Tạo thêm:
```text
velocity = p_t - p_{t-1}
acceleration = v_t - v_{t-1}
torso angle
hip drop
body aspect ratio
center-of-body velocity
```
Fall thường có:
```text
torso angle thay đổi nhanh
hip y-coordinate tụt nhanh
bbox height/width thay đổi
center-of-body velocity lớn
```
Đây là điểm giúp model chuyên biệt cho fall detection.
---
# 11. Mô hình DynaFall-GCN
## Kiến trúc đề xuất
```text
Input skeleton sequence
↓
Confidence-aware keypoint dropout
↓
Joint GCN branch
↓
Bone GCN branch
↓
Dynamics temporal branch
↓
Feature fusion
↓
MLP classifier
↓
Fall / Non-fall
```
## Chi tiết 3 branch
### Branch 1: Joint GCN
Học quan hệ giữa các khớp.
```text
Input: normalized joint coordinates
Encoder: lightweight ST-GCN / graph-temporal block
```
### Branch 2: Bone GCN
Học hướng và độ thay đổi của xương.
```text
Input: bone vectors
Encoder: graph-temporal block
```
### Branch 3: Dynamics branch
Học các đặc trưng rơi.
```text
Input: velocity, acceleration, torso angle, hip drop, aspect ratio
Encoder: TCN + attention pooling
```
### Fusion
```text
f = concat(f_joint, f_bone, f_dyn)
logits = MLP(f)
```
---
# 12. Confidence-aware keypoint dropout
Đây là điểm mới dễ viết.
## Ý tưởng
Trong thực tế, pose estimator có thể mất keypoint do:
```text
che khuất
camera góc xấu
người nằm sát sàn
motion blur
```
Do đó, khi train, mình chủ động mask keypoints.
## Cách làm
Nếu confidence thấp:
```text
conf < 0.3
```
thì xác suất bị mask cao hơn.
Ví dụ:
```text
high-confidence keypoint: mask probability = 0.1
low-confidence keypoint: mask probability = 0.5
```
Khi mask:
```text
x = 0
y = 0
confidence = 0
velocity = 0
acceleration = 0
```
---
# 13. Thí nghiệm cần chạy
## Experiment 1: Main comparison
Chạy tất cả method trên từng dataset.
```text
URFD train/test
MCFD train/test
```
Bảng:
| Method | URFD Acc | URFD Recall | URFD F1 | MCFD Acc | MCFD Recall | MCFD F1 |
| ---------------- | -------: | ----------: | ------: | -------: | ----------: | ------: |
| LSTM | | | | | | |
| ST-GCN | | | | | | |
| 2s-AGCN | | | | | | |
| CTR-GCN | | | | | | |
| PoseC3D | | | | | | |
| TCNTE-style | | | | | | |
| **DynaFall-GCN** | | | | | | |
---
## Experiment 2: Ablation study
Chỉ chạy trên model của mình.
| Variant | Joint | Bone | Dynamics | Dropout | F1 | Recall |
| ------- | ----- | ---- | -------- | ------------------------ | -: | -----: |
| A1 | ✓ | | | | | |
| A2 | ✓ | ✓ | | | | |
| A3 | ✓ | ✓ | ✓ | | | |
| A4 | ✓ | ✓ | ✓ | Random dropout | | |
| A5 | ✓ | ✓ | ✓ | Confidence-aware dropout | | |
Mục tiêu:
```text
Bone giúp tăng F1.
Dynamics giúp tăng Recall.
Confidence-aware dropout giúp robustness tốt hơn.
```
---
## Experiment 3: Robustness under missing keypoints
Tạo test set bị mất keypoint:
```text
Clean
Missing-10
Missing-20
Missing-30
Upper-body missing
Lower-body missing
Low-confidence masking
```
Bảng:
| Method | Clean F1 | Missing-10 | Missing-20 | Missing-30 | Lower-body missing |
| ---------------- | -------: | ---------: | ---------: | ---------: | -----------------: |
| ST-GCN | | | | | |
| 2s-AGCN | | | | | |
| CTR-GCN | | | | | |
| PoseC3D | | | | | |
| TCNTE-style | | | | | |
| **DynaFall-GCN** | | | | | |
Đây là bảng nên nhấn mạnh trong paper.
---
## Experiment 4: Cross-dataset generalization
```text
Train URFD → Test MCFD
Train MCFD → Test URFD
```
Bảng:
| Train | Test | ST-GCN | 2s-AGCN | CTR-GCN | PoseC3D | TCNTE-style | Ours |
| ----- | ---- | -----: | ------: | ------: | ------: | ----------: | ---: |
| URFD | MCFD | | | | | | |
| MCFD | URFD | | | | | | |
Nếu kết quả thấp cũng không sao. Cross-dataset thường khó. Chỉ cần proposed giảm ít hơn baseline.
---
## Experiment 5: Runtime / Efficiency
Báo cáo:
```text
Parameters
FLOPs nếu tính được
Inference time per clip
FPS
```
Bảng:
| Method | Params | FPS | F1 |
| ------------ | -----: | --: | -: |
| ST-GCN | | | |
| CTR-GCN | | | |
| PoseC3D | | | |
| TCNTE-style | | | |
| DynaFall-GCN | | | |
---
# 14. Metrics bắt buộc
Dùng:
```text
Accuracy
Precision
Recall / Sensitivity
Specificity
F1-score
Macro-F1
FPS
```
Trong fall detection, cần nhấn mạnh:
```text
Recall quan trọng hơn Accuracy vì bỏ sót fall nguy hiểm hơn báo nhầm.
```
---
# 15. Training setup
## Split
Nếu dataset nhỏ:
```text
5-fold cross-validation
```
Nếu muốn nhanh:
```text
70% train
15% validation
15% test
```
Nhưng phải là:
```text
video-level split
```
không phải frame-level split.
## Seeds
Tối thiểu:
```text
7, 13, 21
```
Tốt hơn:
```text
7, 13, 21, 42, 100
```
## Hyperparameters gợi ý
| Setting | Value |
| -------------- | ---------------------------------: |
| Clip length | 32 |
| Stride | 16 |
| Batch size | 16 hoặc 32 |
| Epochs | 80–100 |
| Optimizer | AdamW |
| Learning rate | 1e-3 hoặc 5e-4 |
| Weight decay | 1e-4 |
| Early stopping | patience 10 |
| Loss | CrossEntropy / Weighted Focal Loss |
---
# 16. Cấu trúc paper 8 trang
## Title
**DynaFall: Robust Skeleton-Based Fall Detection via Keypoint Dynamics and Graph-Based Benchmarking**
## Abstract
Nêu 5 ý:
```text
fall detection quan trọng
RGB-based methods có privacy/computation issue
skeleton-based methods nhẹ hơn nhưng dễ bị pose noise
paper benchmark các skeleton methods mạnh
đề xuất DynaFall-GCN với dynamics + confidence-aware dropout
```
## 1. Introduction
Flow:
```text
Fall detection quan trọng cho elderly care
Wearable sensors bất tiện, RGB video có privacy concern
Skeleton/keypoints là hướng nhẹ và privacy-preserving
Gap: nhiều bài chỉ so baseline đơn giản, ít benchmark method GCN/PoseC3D mạnh trên fall datasets
Gap: ít kiểm thử robustness khi mất keypoint
Contribution
```
## 2. Related Work
Chia 4 phần:
```text
2.1 Vision-based fall detection
2.2 Skeleton-based action recognition
2.3 Graph-based skeleton models
2.4 Robust fall detection under occlusion
```
## 3. Proposed Method
```text
3.1 Overview
3.2 Pose extraction
3.3 Skeleton normalization
3.4 Joint, bone, and dynamics representation
3.5 Confidence-aware keypoint dropout
3.6 DynaFall-GCN architecture
```
## 4. Experiments
```text
4.1 Datasets
4.2 Comparative methods
4.3 Evaluation protocol
4.4 Implementation details
4.5 Metrics
```
## 5. Results and Discussion
```text
5.1 Main comparison
5.2 Ablation study
5.3 Robustness analysis
5.4 Cross-dataset generalization
5.5 Runtime analysis
5.6 Error analysis
```
## 6. Conclusion
Nêu:
```text
DynaFall-GCN hiệu quả và robust
skeleton giúp privacy-preserving
hạn chế: dataset vẫn là simulated falls
future work: real-world elderly care, edge deployment, multimodal fusion
```
---
# 17. Figures nên có
## Figure 1: Overall framework
```text
RGB Video
→ Pose Extraction
→ Skeleton Normalization
→ Joint/Bone/Dynamics Features
→ Comparative Skeleton Methods
→ DynaFall-GCN
→ Fall / Non-fall
```
## Figure 2: DynaFall-GCN architecture
```text
Joint Stream
Bone Stream
Dynamics Stream
→ Fusion
→ Classifier
```
## Figure 3: Robustness scenarios
```text
Clean skeleton
Missing-10
Missing-20
Missing-30
Lower-body occlusion
```
---
# 18. Bảng cần có trong paper
Tối thiểu nên có 5 bảng:
```text
Table 1. Dataset statistics
Table 2. Main comparison with related methods
Table 3. Ablation study
Table 4. Robustness under missing keypoints
Table 5. Cross-dataset generalization
```
Nếu thiếu trang thì bỏ Table 5 hoặc đưa vào phụ lục, nhưng **không nên bỏ Table 3 và Table 4**.
---
# 19. Lộ trình chạy thực nghiệm
## Phase 1: Dataset + pose extraction
Làm:
```text
Download URFD + MCFD
Extract frames
Run YOLO pose
Save keypoints
Visual check skeleton
```
Output:
```text
URFD_keypoints.pkl
MCFD_keypoints.pkl
```
---
## Phase 2: Build clips
Làm:
```text
video-level split
32-frame clips
stride 16
normalize skeleton
generate joint/bone/dynamics features
```
Output:
```text
URFD_train.pkl / val.pkl / test.pkl
MCFD_train.pkl / val.pkl / test.pkl
```
---
## Phase 3: Chạy baseline dễ trước
Chạy:
```text
LSTM
ST-GCN
```
Mục tiêu:
```text
kiểm tra data pipeline đúng chưa
```
---
## Phase 4: Chạy method từ paper khác
Chạy:
```text
2s-AGCN
CTR-GCN
PoseC3D
TCNTE-style
```
---
## Phase 5: Chạy proposed
Chạy:
```text
DynaFall-GCN joint only
DynaFall-GCN joint + bone
DynaFall-GCN joint + bone + dynamics
DynaFall-GCN full + confidence-aware dropout
```
---
## Phase 6: Robustness test
Test tất cả model chính trên:
```text
Clean
Missing-10
Missing-20
Missing-30
Lower-body missing
Low-confidence masking
```
---
## Phase 7: Viết paper
Thứ tự viết nên là:
```text
Method
Experiments
Results
Introduction
Related Work
Abstract
Conclusion
```
---
# 20. Bản tối giản nếu thiếu thời gian
Nếu deadline gấp, chỉ làm:
```text
Datasets:
URFD + MCFD
Methods:
LSTM
ST-GCN
2s-AGCN
CTR-GCN
TCNTE-style
DynaFall-GCN
Bỏ nếu không kịp:
PoseC3D
Cross-dataset
5 seeds
```
Nhưng giữ bắt buộc:
```text
Main comparison
Ablation
Robustness missing keypoints
```
---
# 21. Kế hoạch cuối cùng nên theo
Chốt hướng như sau:
```text
Paper title:
DynaFall: Robust Skeleton-Based Fall Detection via Keypoint Dynamics and Graph-Based Benchmarking
Datasets:
URFD + MCFD
UP-Fall optional
Pose extractor:
YOLOv8n-pose hoặc YOLO11n-pose
Compared methods:
LSTM
ST-GCN
2s-AGCN
CTR-GCN
PoseC3D nếu kịp
TCNTE-style
Proposed:
DynaFall-GCN
= joint stream
+ bone stream
+ fall-specific dynamics stream
+ confidence-aware keypoint dropout
Experiments:
Main results
Ablation
Robustness
Cross-dataset nếu kịp
Runtime/FPS
Main claim:
DynaFall-GCN is a lightweight and privacy-preserving skeleton-based fall detection model that achieves competitive performance and better robustness under missing or unreliable keypoints.
```
Đây là kế hoạch hợp lý nhất cho hội nghị Q4: **đủ mới, đủ method từ paper khác, đủ thí nghiệm, nhưng vẫn chạy được trong thời gian ngắn**.
[1]: https://fenix.ur.edu.pl/~mkepski/ds/uf.html?utm_source=chatgpt.com "UR Fall Detection Dataset"
[2]: https://www.iro.umontreal.ca/~labimage/Dataset/?utm_source=chatgpt.com "Multiple cameras fall dataset"
[3]: https://sites.google.com/up.edu.mx/challenge-up-2019/data?utm_source=chatgpt.com "Challenge UP - Data"
[4]: https://www.mdpi.com/1424-8220/19/9/1988?utm_source=chatgpt.com "UP-Fall Detection Dataset: A Multimodal Approach"
[5]: https://ojs.aaai.org/index.php/aaai/article/view/12328?utm_source=chatgpt.com "Spatial Temporal Graph Convolutional Networks ..."
[6]: https://arxiv.org/abs/1805.07694?utm_source=chatgpt.com "Two-Stream Adaptive Graph Convolutional Networks for Skeleton-Based Action Recognition"
[7]: https://arxiv.org/abs/2107.12213?utm_source=chatgpt.com "Channel-wise Topology Refinement Graph Convolution for Skeleton-Based Action Recognition"
[8]: https://openaccess.thecvf.com/content/CVPR2022/html/Duan_Revisiting_Skeleton-Based_Action_Recognition_CVPR_2022_paper.html?utm_source=chatgpt.com "Related Material - CVPR 2022 Open Access Repository"
[9]: https://www.sciencedirect.com/science/article/abs/pii/S1574119225000057?utm_source=chatgpt.com "A Real-time skeleton-based fall detection algorithm ..."
|