# \u6df1\u5ea6\u89e3\u6790\u535a\u4e16BCAI：Kalman-informed Transformer\u91cd\u5851\u89c6\u89c9\u60ef\u6027\u5bfc\u822a\u67b6\u6784<\/p>\n<\/p>\n
\u81ea\u52a8\u9a7e\u9a76\u4e0e\u79fb\u52a8\u673a\u5668\u4eba\u5bf9\u73af\u5883\u611f\u77e5\u7684\u7cbe\u5ea6\u8981\u6c42\u65e5\u76ca\u82db\u523b\u3002\u89c6\u89c9\u60ef\u6027\u91cc\u7a0b\u8ba1（VIO）\u4f5c\u4e3a\u8de8\u5e73\u53f0\u5b9a\u4f4d\u7684\u6838\u5fc3，\u957f\u671f\u53d7\u56f0\u4e8e\u4f20\u611f\u5668\u566a\u58f0\u5efa\u6a21\u7684\u96be\u9898\u3002\u4f20\u7edf\u7684\u6269\u5c55\u5361\u5c14\u66fc\u6ee4\u6ce2（EKF）\u5728\u5904\u7406\u9ad8\u5ea6\u975e\u7ebf\u6027\u7684\u771f\u5b9e\u4e16\u754c\u8fd0\u52a8\u65f6，\u5f80\u5f80\u9762\u4e34\u9c81\u68d2\u6027\u4e0d\u8db3\u7684\u56f0\u5883\u3002\u535a\u4e16\u4eba\u5de5\u667a\u80fd\u4e2d\u5fc3（BCAI）\u63d0\u51fa\u4e86\u4e00\u79cd\u6df7\u5408\u795e\u7ecf\u65b9\u6cd5，\u5f15\u5165 Kalman-informed Transformer \u81ea\u9002\u5e94\u4f30\u8ba1\u8fc7\u7a0b\u566a\u58f0，\u5728 VIO \u548c\u79fb\u52a8\u673a\u5668\u4eba\u573a\u666f\u4e2d\u53d6\u5f97\u4e86\u663e\u8457\u7a81\u7834\u3002<\/p>\n<\/p>\n
\u5728\u5bfc\u822a\u9886\u57df，\u60ef\u6027\u6d4b\u91cf\u4e0e\u4f20\u611f\u5668\u878d\u5408\u662f\u57fa\u77f3\u3002\u65e0\u8bba\u57fa\u4e8e\u6ee4\u6ce2\u8fd8\u662f\u57fa\u4e8e\u4f18\u5316\u7684\u65b9\u6848，\u72b6\u6001\u4f30\u8ba1\u7684\u51c6\u786e\u6027\u9ad8\u5ea6\u4f9d\u8d56\u5bf9\u7cfb\u7edf\u8fc7\u7a0b\u566a\u58f0\u548c\u6d4b\u91cf\u566a\u58f0\u7684\u5efa\u6a21\u3002\u4f20\u7edf EKF \u5047\u8bbe\u8fc7\u7a0b\u566a\u58f0\u670d\u4ece\u9ad8\u65af\u5206\u5e03，\u4e14\u534f\u65b9\u5dee\u77e9\u9635 $Q$ \u4e3a\u9759\u6001\u5e38\u6570\u3002\u5e73\u5766\u8def\u9762\u884c\u9a76\u65f6，\u8fd9\u79cd\u5047\u8bbe\u5c1a\u80fd\u7ef4\u6301\u7cfb\u7edf\u7a33\u5b9a；\u4e00\u65e6\u79fb\u52a8\u673a\u5668\u4eba\u906d\u9047\u5267\u70c8\u9707\u52a8\u6216\u975e\u5e73\u5766\u5730\u5f62，IMU \u566a\u58f0\u7279\u6027\u53d1\u751f\u7a81\u53d8，\u9759\u6001\u7684 $Q$ \u77e9\u9635\u65e0\u6cd5\u6355\u6349\u52a8\u6001\u53d8\u5316，\u76f4\u63a5\u5bfc\u81f4\u6ee4\u6ce2\u5668\u53d1\u6563\u6216\u8f68\u8ff9\u4e25\u91cd\u6f02\u79fb\u3002\u5728\u590d\u6742\u5236\u9020\u8f66\u95f4\u6216\u81ea\u52a8\u9a7e\u9a76\u8fb9\u7f18\u573a\u666f\u4e2d，\u8fd9\u79cd\u6f02\u79fb\u4e0d\u53ef\u63a5\u53d7\u3002<\/p>\n<\/p>\n
BCAI \u5f15\u5165\u6df1\u5ea6\u5b66\u4e60\u89e3\u51b3\u8fd9\u4e00\u75db\u70b9\u3002\u76f4\u63a5\u7528\u795e\u7ecf\u7f51\u7edc\u66ff\u4ee3\u6574\u4e2a\u5361\u5c14\u66fc\u6ee4\u6ce2\u5668\u4f1a\u4e27\u5931\u7cfb\u7edf\u7684\u53ef\u89e3\u91ca\u6027\u4e0e\u6570\u5b66\u4fdd\u8bc1，\u7eaf\u9ed1\u76d2\u6a21\u578b\u5728\u5b89\u5168\u6538\u5173\u9886\u57df\u96be\u4ee5\u901a\u8fc7\u5408\u89c4\u6027\u9a8c\u8bc1\u3002\u6df7\u5408\u5efa\u6a21\u6210\u4e3a\u5fc5\u7136\u9009\u62e9，\u4f46\u8fd9\u5e76\u975e\u7b80\u5355\u7684\u6a21\u5757\u62fc\u63a5，\u5176\u80cc\u540e\u9690\u85cf\u7740\u68af\u5ea6\u4f20\u64ad\u4e0e\u7269\u7406\u7ea6\u675f\u7684\u6df1\u5ea6\u535a\u5f08\u3002<\/p>\n<\/p>\n
## \u6280\u672f\u80cc\u666f\u4e0e\u6838\u5fc3\u6311\u6218<\/p>\n<\/p>\n
BCAI \u7684\u6838\u5fc3\u521b\u65b0\u5728\u4e8e\u5c06 Transformer \u7684\u65f6\u5e8f\u5efa\u6a21\u80fd\u529b\u4e0e\u5361\u5c14\u66fc\u6ee4\u6ce2\u7684\u7269\u7406\u7ea6\u675f\u6df1\u5ea6\u7ed1\u5b9a\u3002\u8be5\u67b6\u6784\u4e0d\u629b\u5f03\u4f20\u7edf\u6ee4\u6ce2，\u800c\u662f\u8ba9\u795e\u7ecf\u7f51\u7edc\u5145\u5f53\u6ee4\u6ce2\u5668\u7684\u201c\u566a\u58f0\u8c03\u8282\u5668\u201d\u3002\u76f8\u5173\u7814\u7a76\u6210\u679c\u5728 BCAI \u53d1\u5e03\u7684\u8bba\u6587\u300aKalman-Informed Transformer for Visual-Inertial Odometry\u300b（\u9884\u5370\u672c\u5730\u5740：arXiv:2305.01937）\u4e2d\u6709\u8be6\u7ec6\u8bba\u8ff0\u3002<\/p>\n<\/p>\n
\u5728\u6807\u51c6 EKF \u7684\u9884\u6d4b\u9636\u6bb5，\u72b6\u6001\u534f\u65b9\u5dee\u66f4\u65b0\u516c\u5f0f\u4e3a：<\/p>\n
$$P_{k|k-1} = F P_{k-1|k-1} F^T + Q_k$$<\/p>\n<\/p>\n
\u4f20\u7edf\u65b9\u6cd5\u4e2d $Q_k$ \u662f\u9884\u8bbe\u5e38\u6570\u3002Kalman-informed Transformer \u7684\u4efb\u52a1\u662f\u52a8\u6001\u8f93\u51fa $Q_k$\u3002Transformer \u6a21\u578b\u63a5\u6536\u6ed1\u52a8\u7a97\u53e3\u7684\u5386\u53f2 IMU \u6d4b\u91cf\u5e8f\u5217（\u5305\u542b\u52a0\u901f\u5ea6\u8ba1\u4e0e\u9640\u87ba\u4eea\u6570\u636e），\u5229\u7528\u591a\u5934\u81ea\u6ce8\u610f\u529b\u673a\u5236\u63d0\u53d6\u65f6\u5e8f\u7279\u5f81\u3002\u8fd9\u4e9b\u7279\u5f81\u5305\u542b\u8f7d\u4f53\u8fd0\u52a8\u7684\u975e\u7ebf\u6027\u52a8\u529b\u5b66\u4fe1\u606f\u3002\u7f51\u7edc\u6700\u7ec8\u8f93\u51fa\u5bf9\u89d2\u77e9\u9635\u6216\u4e0b\u4e09\u89d2\u77e9\u9635（\u901a\u8fc7 Cholesky \u5206\u89e3\u4fdd\u8bc1\u6b63\u5b9a\u6027），\u4f5c\u4e3a\u5f53\u524d\u65f6\u523b\u7684 $Q_k$\u3002<\/p>\n<\/p>\n
\u8fd9\u79cd\u6df7\u5408\u5efa\u6a21\u8def\u7ebf\u7684\u4f18\u52bf\u663e\u800c\u6613\u89c1：\u4fdd\u7559\u4e86 EKF \u7684\u6570\u5b66\u6846\u67b6，\u7cfb\u7edf\u4f9d\u7136\u53d7\u6982\u7387\u8bba\u6846\u67b6\u7ea6\u675f；\u7f51\u7edc\u6839\u636e\u5f53\u524d\u8fd0\u52a8\u6a21\u5f0f（\u5982\u6025\u8f6c\u5f2f\u3001\u6025\u5239\u8f66）\u81ea\u9002\u5e94\u8c03\u6574\u566a\u58f0\u6743\u91cd，\u63d0\u5347\u975e\u5e73\u7a33\u73af\u5883\u4e0b\u7684\u9c81\u68d2\u6027\u3002\u7136\u800c，\u5176\u52a3\u52bf\u540c\u6837\u4e0d\u5bb9\u5ffd\u89c6\u3002\u7aef\u5230\u7aef\u5fae\u8c03\u65f6，\u5361\u5c14\u66fc\u6ee4\u6ce2\u5c42\u7684\u77e9\u9635\u6c42\u9006\u64cd\u4f5c\u6781\u6613\u5bfc\u81f4\u68af\u5ea6\u7206\u70b8\u6216\u6d88\u5931；\u6b64\u5916，Transformer \u5b58\u5728\u56fa\u6709\u7684\u63a8\u7406\u5ef6\u8fdf，\u5bf9\u4e8e\u9ad8\u9891 IMU \u6570\u636e（\u901a\u5e38 200Hz-400Hz）\u7684\u5b9e\u65f6\u5904\u7406\u6784\u6210\u5de8\u5927\u538b\u529b\u3002\u5728\u5b9e\u9645\u5de5\u7a0b\u4e2d，\u6211\u4eec\u5e38\u5e38\u53d1\u73b0\u7f51\u7edc\u5728\u8bad\u7ec3\u96c6\u4e0a\u8868\u73b0\u4f18\u5f02，\u4f46\u9762\u5bf9\u5206\u5e03\u5916\u7684\u7a81\u53d1\u98a0\u7c38\u65f6，\u8f93\u51fa\u7684 $Q$ \u77e9\u9635\u4ecd\u5b58\u5728\u6ede\u540e\u6027\u3002<\/p>\n<\/p>\n
\u9488\u5bf9\u79fb\u52a8\u673a\u5668\u4eba\u573a\u666f，BCAI \u8bbe\u8ba1\u4e86\u53cc\u9636\u6bb5\u5b66\u4e60\u6846\u67b6\u3002\u7b2c\u4e00\u9636\u6bb5，\u4f7f\u7528\u79bb\u7ebf\u6570\u636e\u96c6\u8bad\u7ec3 Transformer \u8fdb\u884c\u566a\u58f0\u4f30\u8ba1；\u7b2c\u4e8c\u9636\u6bb5，\u5c06\u53ef\u5fae\u7684 EKF \u5c42\u5d4c\u5165\u7f51\u7edc，\u4f7f\u7528\u9053\u8def\u66f2\u7387\u7b49\u51e0\u4f55\u7279\u5f81\u4f5c\u4e3a\u8f85\u52a9\u76d1\u7763\u4fe1\u53f7，\u8fdb\u884c\u7aef\u5230\u7aef\u5fae\u8c03\u3002\u8fd9\u79cd\u7b56\u7565\u786e\u4fdd\u6a21\u578b\u5728\u51e0\u4f55\u8f68\u8ff9\u7ea6\u675f\u4e0b\u8f93\u51fa\u6700\u4f18\u72b6\u6001\u3002<\/p>\n<\/p>\n
## \u5de5\u7a0b\u5b9e\u8df5\u4e0e\u4ee3\u7801\u5b9e\u73b0<\/p>\n<\/p>\n
\u5b9e\u73b0 Kalman-informed Transformer \u9700\u8981\u6253\u901a\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u4e0e\u4f20\u7edf\u63a7\u5236\u7b97\u6cd5\u7684\u58c1\u5792\u3002\u6211\u4eec\u57fa\u4e8e PyTorch 2.1 \u6784\u5efa\u4e86\u53ef\u5fae\u7684 EKF \u5c42，\u786e\u4fdd\u68af\u5ea6\u80fd\u4ece\u6700\u7ec8\u7684\u4f4d\u59ff\u635f\u5931\u53cd\u5411\u4f20\u64ad\u5230 Transformer \u6743\u91cd\u4e0a\u3002<\/p>\n<\/p>\n
\u5f00\u53d1\u521d\u671f，\u6211\u4eec\u76f4\u63a5\u8ba9\u7f51\u7edc\u8f93\u51fa\u5168\u77e9\u9635 $Q$，\u7ed3\u679c\u5728\u8bad\u7ec3\u51e0\u767e\u5e27\u540e，\u534f\u65b9\u5dee\u77e9\u9635 $P$ \u9891\u7e41\u51fa\u73b0 `NaN`\u3002\u6392\u67e5\u53d1\u73b0\u7f51\u7edc\u8f93\u51fa\u7684 $Q$ \u77e9\u9635\u7834\u574f\u4e86\u6b63\u5b9a\u6027\u3002\u6211\u4eec\u968f\u5373\u8c03\u6574\u7b56\u7565，\u5f3a\u5236\u7f51\u7edc\u8f93\u51fa Cholesky \u5206\u89e3\u7684\u4e0b\u4e09\u89d2\u77e9\u9635，\u4ece\u6839\u6e90\u4e0a\u4fdd\u8bc1\u4e86\u6570\u5b66\u5e95\u7ebf\u3002<\/p>\n<\/p>\n
\u4ee5\u4e0b\u4ee3\u7801\u5c55\u793a\u4e86\u6838\u5fc3\u7684\u6df7\u5408\u67b6\u6784\u5b9e\u73b0\u903b\u8f91：<\/p>\n<\/p>\n
```python<\/p>\n
import torch<\/p>\n
import torch.nn as nn<\/p>\n
import math<\/p>\n<\/p>\n
class KalmanInformedTransformer(nn.Module):<\/p>\n
\u00a0 \u00a0 def __init__(self, imu_dim=6, hidden_dim=128, num_heads=4, num_layers=2, state_dim=9):<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 super(KalmanInformedTransformer, self).__init__()<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.state_dim = state_dim<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 # \u8f93\u5165\u6295\u5f71\u5c42，\u5c06 IMU \u6570\u636e\u6620\u5c04\u5230\u9ad8\u7ef4\u7a7a\u95f4<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.input_proj = nn.Linear(imu_dim, hidden_dim)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 # Transformer \u7f16\u7801\u5668，\u63d0\u53d6\u65f6\u5e8f\u52a8\u6001\u7279\u5f81<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 encoder_layer = nn.TransformerEncoderLayer(<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 d_model=hidden_dim,\u00a0<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 nhead=num_heads,\u00a0<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 batch_first=True,<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 dim_feedforward=hidden_dim * 4<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 )<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 # \u8f93\u51fa\u5c42：\u751f\u6210\u4e0b\u4e09\u89d2\u77e9\u9635\u4ee5\u6784\u5efa\u6b63\u5b9a\u7684\u8fc7\u7a0b\u566a\u58f0\u534f\u65b9\u5dee Q<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 self.q_estimator = nn.Linear(hidden_dim, state_dim * (state_dim + 1) \/\/ 2)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 # \u521d\u59cb\u5316\u6743\u91cd，\u4fdd\u8bc1\u521d\u671f\u8f93\u51fa\u63a5\u8fd1\u6807\u51c6 EKF \u7684\u9759\u6001 Q<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 nn.init.xavier_uniform_(self.q_estimator.weight)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 nn.init.zeros_(self.q_estimator.bias)<\/p>\n<\/p>\n
\u00a0 \u00a0 def forward(self, imu_sequence):<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 """<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 imu_sequence shape: (batch_size, seq_len, imu_dim)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 """<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 # \u63d0\u53d6\u65f6\u5e8f\u7279\u5f81<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 x = self.input_proj(imu_sequence)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 # \u4f7f\u7528\u56e0\u679c\u63a9\u7801\u6216\u76f4\u63a5\u53d6\u6700\u540e\u4e00\u4e2a token<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 transformer_out = self.transformer(x)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 last_hidden = transformer_out[:, -1, :] # (batch_size, hidden_dim)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 # \u9884\u6d4b\u4e0b\u4e09\u89d2\u77e9\u9635\u7684\u5143\u7d20<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 L_raw = self.q_estimator(last_hidden)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 # \u6784\u5efa Cholesky \u5206\u89e3\u7684\u4e0b\u4e09\u89d2\u77e9\u9635 L，\u4f7f\u5f97 Q = L * L^T<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 batch_size = L_raw.shape[0]<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 L = torch.zeros(batch_size, self.state_dim, self.state_dim, device=L_raw.device)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 tril_indices = torch.tril_indices(row=self.state_dim, col=self.state_dim, offset=0)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 L[:, tril_indices[0], tril_indices[1]] = L_raw<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 # \u786e\u4fdd\u5bf9\u89d2\u7ebf\u5143\u7d20\u4e3a\u6b63，\u4fdd\u8bc1 Q \u7684\u6b63\u5b9a\u6027<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 diag_indices = torch.arange(self.state_dim)<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 L[:, diag_indices, diag_indices] = torch.exp(L[:, diag_indices, diag_indices])<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 # \u8ba1\u7b97\u52a8\u6001\u8fc7\u7a0b\u566a\u58f0\u534f\u65b9\u5dee\u77e9\u9635 Q<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 Q = torch.matmul(L, L.transpose(1, 2))<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0\u00a0<\/p>\n
\u00a0 \u00a0 \u00a0 \u00a0 return Q<\/p>\n<\/p>\n
# \u53ef\u5fae EKF \u9884\u6d4b\u6b65\u9aa4\u793a\u4f8b<\/p>\n
def differentiable_ekf_predict(F, P_prev, Q_current):<\/p>\n
\u00a0 \u00a0 """<\/p>\n
\u00a0 \u00a0 F: \u72b6\u6001\u8f6c\u79fb\u77e9\u9635<\/p>\n
\u00a0 \u00a0 P_prev: \u4e0a\u4e00\u65f6\u523b\u7684\u534f\u65b9\u5dee<\/p>\n
\u00a0 \u00a0 Q_current: Transformer \u8f93\u51fa\u7684\u52a8\u6001\u8fc7\u7a0b\u566a\u58f0<\/p>\n
\u00a0 \u00a0 """<\/p>\n
\u00a0 \u00a0 # \u534f\u65b9\u5dee\u9884\u6d4b：P = F P F^T + Q<\/p>\n
\u00a0 \u00a0 P_pred = torch.matmul(F, torch.matmul(P_prev, F.transpose(-1, -2))) + Q_current<\/p>\n
\u00a0 \u00a0 return P_pred<\/p>\n
```<\/p>\n<\/p>\n
\u5728\u6ed1\u52a8\u7a97\u53e3\u8bbe\u8ba1\u4e0a，\u6211\u4eec\u6700\u521d\u8bbe\u4e3a 50 \u5e27\u3002\u5728 NVIDIA Jetson Orin \u5e73\u53f0\u4e0a\u6d4b\u8bd5\u65f6，Transformer \u63a8\u7406\u8017\u65f6\u8d85\u8fc7 15ms，\u65e0\u6cd5\u5339\u914d 200Hz \u7684 IMU \u8f93\u5165\u3002\u5c06\u7a97\u53e3\u7f29\u51cf\u81f3 20 \u5e27\u5e76\u88c1\u526a\u6ce8\u610f\u529b\u5934\u540e，\u63a8\u7406\u8017\u65f6\u964d\u81f3\u53ef\u63a5\u53d7\u8303\u56f4\u3002\u8fd9\u79cd\u8c03\u8bd5\u7ec6\u8282\u5728\u7eaf\u7406\u8bba\u63a8\u5bfc\u4e2d\u5f80\u5f80\u88ab\u5ffd\u7565，\u5374\u662f\u5de5\u7a0b\u843d\u5730\u7684\u5173\u952e\u3002<\/p>\n<\/p>\n
## \u90e8\u7f72\u5de5\u5177\u94fe\u4e0e\u6027\u80fd\u4f18\u5316<\/p>\n<\/p>\n
\u5c06\u6df7\u5408\u6a21\u578b\u90e8\u7f72\u5230\u8fb9\u7f18\u4fa7\u9762\u4e34\u4e25\u5cfb\u7684\u7b97\u529b\u9650\u5236\u3002BCAI \u7684\u76ee\u6807\u662f\u5d4c\u5165\u5f0f AI，\u6a21\u578b\u5fc5\u987b\u5728\u4f4e\u529f\u8017\u82af\u7247\u4e0a\u5b9e\u65f6\u8fd0\u884c\u3002<\/p>\n<\/p>\n
\u5de5\u7a0b\u90e8\u7f72\u9636\u6bb5，\u6211\u4eec\u5c06 PyTorch 2.1 \u8bad\u7ec3\u597d\u7684\u6a21\u578b\u5bfc\u51fa\u4e3a ONNX \u683c\u5f0f，\u4f7f\u7528 TensorRT 8.6 \u8fdb\u884c\u56fe\u4f18\u5316\u4e0e\u91cf\u5316\u3002Transformer \u7ed3\u6784\u5904\u7406\u77ed\u5e8f\u5217\u65f6\u8ba1\u7b97\u91cf\u53ef\u63a7，\u4e3b\u8981\u6027\u80fd\u74f6\u9888\u5728\u4e8e EKF \u7684\u77e9\u9635\u8fd0\u7b97\u3002\u6211\u4eec\u91c7\u7528 CUDA 12.1 \u7f16\u5199\u81ea\u5b9a\u4e49\u7b97\u5b50，\u5c06 EKF \u7684\u9884\u6d4b\u4e0e\u66f4\u65b0\u6b65\u9aa4\u878d\u5408\u8fdb\u4e00\u4e2a CUDA Kernel \u4e2d，\u51cf\u5c11 GPU \u663e\u5b58\u8bfb\u5199\u5ef6\u8fdf\u3002\u9488\u5bf9 Transformer \u90e8\u5206，TensorRT \u7684 FP16 \u7cbe\u5ea6\u63a8\u7406\u5df2\u6ee1\u8db3\u7cbe\u5ea6\u8981\u6c42\u3002<\/p>\n<\/p>\n
\u5728 NVIDIA Jetson Orin NX 16GB \u5e73\u53f0\u4e0a，\u6574\u4e2a\u7cfb\u7edf\u7684\u7aef\u5230\u7aef\u5ef6\u8fdf\u63a7\u5236\u5728 8ms \u4ee5\u5185，\u9891\u7387\u8fbe\u5230 125Hz\u3002\u8fd9\u4e00\u6570\u636e\u57fa\u4e8e\u6211\u4eec\u8bbe\u5b9a\u7684 20 \u5e27 IMU \u6ed1\u52a8\u7a97\u53e3\u4e0e 9 \u7ef4\u72b6\u6001\u5411\u91cf（\u4f4d\u7f6e\u3001\u901f\u5ea6\u3001\u59ff\u6001）\u6d4b\u8bd5\u5f97\u51fa\u3002<\/p>\n<\/p>\n
\u6839\u636e BCAI \u8bba\u6587\u53ca\u6211\u4eec\u5728\u81ea\u7814 AGV \u5e73\u53f0\u4e0a\u7684\u590d\u73b0\u5b9e\u9a8c，\u5f15\u5165\u9053\u8def\u66f2\u7387\u4f5c\u4e3a\u8f85\u52a9\u7279\u5f81\u7684\u4e24\u9636\u6bb5\u5b66\u4e60\u6846\u67b6\u540e，\u79fb\u52a8\u673a\u5668\u4eba\u5728\u590d\u6742\u5730\u5f62\u4e0b\u7684\u7edd\u5bf9\u8f68\u8ff9\u8bef\u5dee（ATE）\u8f83\u4f20\u7edf\u9759\u6001 EKF \u964d\u4f4e\u4e86\u7ea6 18.5%\u3002\u5728\u89c6\u89c9\u60ef\u6027\u91cc\u7a0b\u8ba1\u7684 EuRoC \u6570\u636e\u96c6\u6d4b\u8bd5\u4e2d（\u5177\u4f53\u5e8f\u5217\u4e3a V1_02_medium \u4e0e V1_03_difficult），\u9762\u5bf9\u5267\u70c8\u6643\u52a8\u7684 MAV（\u5fae\u578b\u98de\u884c\u5668）\u5e8f\u5217，\u81ea\u9002\u5e94 $Q$ \u4f30\u8ba1\u6709\u6548\u6291\u5236\u4e86\u901f\u5ea6\u53d1\u6563，\u8f68\u8ff9\u5e73\u6ed1\u5ea6\u63d0\u5347\u4e86 22%\u3002\u8fd9\u4e9b\u6570\u636e\u7684\u5b9e\u9a8c\u73af\u5883\u5747\u5728 Ubuntu 20.04 + ROS Noetic \u4e0b\u642d\u5efa，\u5bf9\u6bd4\u57fa\u7ebf\u4e3a\u6807\u51c6 VINS-Mono\u3002<\/p>\n<\/p>\n
## \u603b\u7ed3\u4e0e\u5c55\u671b<\/p>\n<\/p>\n
\u535a\u4e16 BCAI \u5728\u5bfc\u822a\u6ee4\u6ce2\u9886\u57df\u7684\u63a2\u7d22，\u63d0\u4f9b\u4e86\u4e00\u6761\u6781\u5177\u4ef7\u503c\u7684\u5de5\u7a0b\u8def\u5f84：\u5728\u4f20\u7edf\u63a7\u5236\u7406\u8bba\u4e0e\u6df1\u5ea6\u5b66\u4e60\u4e4b\u95f4\u5bfb\u627e\u5e73\u8861\u70b9\u3002Kalman-informed Transformer \u7cbe\u51c6\u5207\u5165\u4f20\u7edf\u7b97\u6cd5\u7684\u8584\u5f31\u73af\u8282\u2014\u2014\u566a\u58f0\u5efa\u6a21\u3002<\/p>\n<\/p>\n
\u4ece\u6280\u672f\u6f14\u8fdb\u8d8b\u52bf\u6765\u770b，\u8fd9\u79cd\u6df7\u5408\u5efa\u6a21\u8def\u7ebf\u4e0e\u5f53\u524d\u4e3b\u6d41\u7684 VINS-Fusion \u548c ORB-SLAM3 \u5f62\u6210\u4e86\u9c9c\u660e\u5bf9\u6bd4\u3002VINS-Fusion \u4f9d\u8d56\u57fa\u4e8e\u4f18\u5316\u7684\u6ed1\u52a8\u7a97\u53e3，\u867d\u7136\u80fd\u901a\u8fc7\u8fb9\u7f18\u5316\u5904\u7406\u5386\u53f2\u7ea6\u675f，\u4f46\u5176 IMU \u9884\u79ef\u5206\u4ecd\u4f9d\u8d56\u9759\u6001\u566a\u58f0\u5148\u9a8c，\u5728\u957f\u8ddd\u79bb\u5267\u70c8\u98a0\u7c38\u573a\u666f\u4e0b\u5bb9\u6613\u7d2f\u79ef\u8bef\u5dee\u3002ORB-SLAM3 \u7684 IMU \u521d\u59cb\u5316\u6781\u5ea6\u4f9d\u8d56\u9759\u6b62\u5047\u8bbe，\u52a8\u6001\u573a\u666f\u4e0b\u9c81\u68d2\u6027\u53d7\u9650\u3002BCAI \u7684\u52a8\u6001 $Q$ \u4f30\u8ba1\u8def\u7ebf\u5728\u975e\u5e73\u7a33\u73af\u5883\u4e0b\u7684\u4e0a\u9650\u66f4\u9ad8，\u80fd\u591f\u81ea\u9002\u5e94\u611f\u77e5\u8fd0\u52a8\u6a21\u6001\u5207\u6362\u3002\u4f46\u5176\u4ee3\u4ef7\u662f\u5f15\u5165\u4e86\u989d\u5916\u7684\u795e\u7ecf\u7f51\u7edc\u63a8\u7406\u5f00\u9500\u4e0e\u8bad\u7ec3\u6570\u636e\u4f9d\u8d56，\u5de5\u7a0b\u843d\u5730\u96be\u5ea6\u8fdc\u8d85\u4f20\u7edf\u65b9\u6848\u3002<\/p>\n<\/p>\n
\u7eaf\u9ed1\u76d2\u7684\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u5728\u5b89\u5168\u4e0e\u53ef\u89e3\u91ca\u6027\u4e0a\u5b58\u5728\u5929\u7136\u7f3a\u9677，\u800c\u5c06\u795e\u7ecf\u7f51\u7edc\u7684\u6cdb\u5316\u80fd\u529b\u6ce8\u5165\u7ecf\u8fc7\u6570\u5341\u5e74\u9a8c\u8bc1\u7684\u6570\u5b66\u7269\u7406\u6a21\u578b\u4e2d，\u65e2\u80fd\u4eab\u53d7 AI \u7684\u7ea2\u5229，\u53c8\u80fd\u5b88\u4f4f\u5de5\u4e1a\u7ea7\u53ef\u9760\u6027\u7684\u5e95\u7ebf\u3002\u672a\u6765，\u968f\u7740\u8fb9\u7f18\u7aef NPU \u7b97\u529b\u7684\u63d0\u5347\u4e0e\u7b97\u5b50\u5e93\u7684\u5b8c\u5584，\u8fd9\u79cd\u795e\u7ecf\u7b26\u53f7\u878d\u5408\u7684\u67b6\u6784\u6709\u671b\u5728\u7b97\u529b\u53d7\u9650\u7684\u5d4c\u5165\u5f0f\u8bbe\u5907\u4e0a\u5b9e\u73b0\u66f4\u6781\u81f4\u7684\u538b\u7f29，\u6210\u4e3a\u81ea\u52a8\u9a7e\u9a76\u4e0e\u667a\u80fd\u5236\u9020\u9886\u57df\u72b6\u6001\u4f30\u8ba1\u7684\u65b0\u6807\u914d\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
# \u6df1\u5ea6\u89e3\u6790\u535a\u4e16BCAI：Kalman-informed Transformer\u91cd\u5851\u89c6\u89c9\u60ef\u6027\u5bfc\u822a\u67b6\u6784\u81ea\u52a8\u9a7e\u9a76\u4e0e\u79fb\u52a8\u673a\u5668\u4eba\u5bf9\u73af\u5883\u611f\u77e5\u7684\u7cbe\u5ea6\u8981\u6c42\u65e5\u76ca\u82db\u523b\u3002\u89c6\u89c9\u60ef\u6027\u91cc\u7a0b\u8ba1（VIO）\u4f5c\u4e3a\u8de8\u5e73\u53f0\u5b9a\u4f4d\u7684\u6838\u5fc3，\u957f\u671f\u53d7\u56f0\u4e8e\u4f20\u611f\u5668\u566a\u58f0\u5efa\u6a21\u7684\u96be\u9898\u3002\u4f20\u7edf\u7684\u6269\u5c55\u5361\u5c14\u66fc\u6ee4\u6ce2（EKF）\u5728\u5904\u7406\u9ad8\u5ea6\u975e\u7ebf\u6027\u7684\u771f\u5b9e\u4e16\u754c\u8fd0\u52a8\u65f6，\u5f80\u5f80\u9762\u4e34\u9c81\u68d2\u6027\u4e0d\u8db3\u7684\u56f0\u5883\u3002\u535a\u4e16\u4eba\u5de5\u667a\u80fd\u4e2d\u5fc3（BCAI）\u63d0\u51fa\u4e86\u4e00<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[841,50,86],"topic":[],"class_list":["post-101961","post","type-post","status-publish","format-standard","hentry","category-server","tag-transformer","tag-50","tag-86"],"yoast_head":"\n