{"id":5565,"date":"2023-06-07T10:08:50","date_gmt":"2023-06-07T02:08:50","guid":{"rendered":"http:\/\/192.168.10.115\/index.php\/2023\/06\/07\/5565\/"},"modified":"2023-06-07T10:15:50","modified_gmt":"2023-06-07T02:15:50","slug":"%e8%87%aa%e5%8a%a8%e9%a9%be%e9%a9%b6-%e8%81%94%e5%90%88%e6%84%9f%e7%9f%a5%e9%a2%84%e6%b5%8b","status":"publish","type":"post","link":"http:\/\/222.128.65.89:18086\/index.php\/2023\/06\/07\/5565\/","title":{"rendered":"2023-06-07 \u81ea\u52a8\u9a7e\u9a76–\u8054\u5408\u611f\u77e5\u9884\u6d4b"},"content":{"rendered":"

1. \u4ec0\u4e48\u662f\u8054\u5408\u611f\u77e5\u9884\u6d4b\uff1f<\/p>\n

1.1 \u81ea\u52a8\u9a7e\u9a76\u7684\u5de5\u4f5c\u6d41\u7a0b<\/p>\n

\u4f20\u7edf\u7684\u81ea\u52a8\u9a7e\u9a76\u65b9\u6cd5\uff0c\u662f\u5206\u6a21\u5757\u8fdb\u884c\u7684 [17]\uff1a<\/p>\n

<\/p>\n

\uff081\uff09\u611f\u77e5\u6a21\u5757\uff1a\u878d\u5408\u4f20\u611f\u5668\u6570\u636e\uff0c\u611f\u77e5\u73af\u5883<\/p>\n

\uff082\uff09\u9884\u6d4b\u6a21\u5757\uff1a\u9884\u6d4b\u7269\u4f53\u6216\u969c\u788d\u7269\u7684\u672a\u6765\u8fd0\u52a8\u8f68\u8ff9<\/p>\n

\uff083\uff09\u89c4\u5212\u6a21\u5757\uff1a\u89c4\u5212\u8def\u5f84<\/p>\n

\uff084\uff09\u63a7\u5236\u6a21\u5757\uff1a\u63a7\u5236\u6c7d\u8f66\u7684\u8fd0\u52a8\uff0c\u8f93\u51fa\u6c7d\u8f66\u8fd0\u884c\u6307\u4ee4<\/p>\n

\u524d\u4e00\u4e2a\u6a21\u5757\u7684\u8f93\u51fa\uff0c\u662f\u540e\u4e00\u4e2a\u6a21\u5757\u7684\u8f93\u5165\u3002\u5f00\u53d1\u7684\u65f6\u5019\u4e5f\u662f\u5206\u522b\u8fdb\u884c\u5f00\u53d1\uff0c\u914d\u7f6e\u4e0d\u540c\u7684\u5de5\u7a0b\u5e08\uff08\u611f\u77e5\u7b97\u6cd5\u5de5\u7a0b\u5e08\u3001\u9884\u6d4b\u7b97\u6cd5\u5de5\u7a0b\u5e08\u3001\u51b3\u7b56\u89c4\u5212\u7b97\u6cd5\u5de5\u7a0b\u5e08\u3002\u3002\u3002\uff09<\/p>\n

\u8fd9\u6837\u5206\u6a21\u5757\u5f00\u53d1\u7684\u597d\u5904\u662f\uff1a\uff081\uff09\u901f\u5ea6\u5feb\uff0c\u80fd\u5e76\u884c\u5f00\u53d1\u6a21\u5757\uff1b\uff082\uff09\u53ef\u89e3\u91ca\u6027\u5f3a\uff0c\u80fd\u77e5\u9053\u6bcf\u4e2a\u6a21\u5757\u7684\u673a\u5236\u3002<\/p>\n

\u4f46\u540c\u65f6\u4e5f\u6709\u4e0d\u5c11\u5f0a\u7aef\uff0c\u6bd4\u5982\uff1a\uff081\uff09\u7ea7\u8054\u5f0f\u5931\u8bef\uff0c\u56e0\u4e3a\u5f00\u53d1\u6bcf\u4e2a\u6a21\u5757\u7684\u65f6\u5019\uff0c\u9ed8\u8ba4\u4f7f\u7528\u524d\u4e00\u4e2a\u6a21\u5757\u7684groung truth\uff1b\u4f46\u5b9e\u9645\u5e94\u7528\u4e0a\u524d\u4e00\u4e2a\u6a21\u5757\u53ef\u80fd\u4f1a\u51fa\u9519\uff0c\u5bfc\u81f4\u540e\u9762\u6a21\u5757\u5f97\u5230\u4e86\u9519\u8bef\u7684\u8f93\u5165\uff0c\u8fdb\u800c\u8f93\u51fa\u9519\u8bef\u7684\u9884\u6d4b\uff0c\u7136\u540e\u53c8\u5bfc\u81f4\u540e\u9762\u7684\u6a21\u5757\u51fa\u9519\u3002\uff082\uff09\u6bcf\u4e2a\u6a21\u5757\u53ea\u80fd\u5f97\u5230\u5c40\u90e8\u6700\u4f18\uff0c\u56e0\u4e3a\u6bcf\u4e2a\u6a21\u5757\u7684\u5355\u72ec\u5f00\u53d1\u7684\uff0c\u800c\u4e0d\u662f\u6839\u636e\u81ea\u5df1\u9a7e\u9a76\u8fd9\u4e2a\u7ec8\u6781\u76ee\u6807\u8fdb\u884c\u4f18\u5316\uff0c\u6240\u4ee5\u53ea\u80fd\u5f97\u5230\u5c40\u90e8\u6700\u4f18\u3002<\/p>\n

1.2 \u65b0\u8d8b\u52bf\uff1a\u591a\u6a21\u5757\u8054\u5408<\/p>\n

\u672a\u6765\u89e3\u51b3\u4e0a\u8ff0\u7f3a\u70b9\uff0c\u5b66\u672f\u754c\u6700\u65b0\u7684\u8d8b\u52bf\u662f\uff0c\u8bb2\u5176\u4e2d\u51e0\u4e2a\u6a21\u5757\u8fdb\u884c\u5408\u5e76\uff0c\u8fdb\u884c\u201c\u7aef\u5230\u7aef\u7684\u5f00\u53d1\uff08end-to-end\uff09\u201d\u3002\u505a\u6cd5\u6709\u5f88\u591a\uff0c\u6bd4\u5982\uff081\uff09\u5f3a\u5316\u5b66\u4e60\u3001\u6a21\u4eff\u5b66\u4e60 [18]\uff1a\u8f93\u5165\u4f20\u611f\u5668\u6570\u636e\uff0c\u76f4\u63a5\u8f93\u51fa\u6c7d\u8f66\u8fd0\u884c\u6307\u4ee4<\/p>\n

<\/p>\n

\uff082\uff09\u7aef\u5230\u7aef\u7684\u8fd0\u52a8\u9884\u6d4b\uff08end-to-end motion prediction\uff09\/ \u8054\u5408\u611f\u77e5\u9884\u6d4b [1,2,3,4,5,6,7,8,13,14,15]\uff1a\u8054\u5408\u5f00\u53d1\u611f\u77e5\u548c\u9884\u6d4b\u6a21\u5757<\/p>\n

\uff083\uff09\u7aef\u5230\u7aef\u7684\u89c4\u5212\uff08end-to-end planning\uff09 [9,11,12,16]\uff1a\u8054\u5408\u5f00\u53d1\u611f\u77e5+\u9884\u6d4b+\u89c4\u5212\u6a21\u5757<\/p>\n

\u8fd9\u79cd\u7aef\u5230\u7aef\u7684\u65b9\u6cd5\u6709\u5f88\u591a\u597d\u5904\uff0c\u6bd4\u5982\u524d\u4e00\u4e2a\u6a21\u5757\u7684\u4fe1\u606f\u53ef\u4ee5\u5e2e\u52a9\u540e\u4e00\u4e2a\u6a21\u5757\u7684\u8ba1\u7b97\u3001\u53ef\u4ee5\u76f4\u63a5\u5229\u7528\u524d\u4e00\u4e2a\u6a21\u5757\u7684\u7279\u5f81\u4ee5\u907f\u514d\u91cd\u590d\u8ba1\u7b97\u3002<\/p>\n

\u5728\u8fd9\u7bc7\u6587\u7ae0\u91cc\uff0c\u6211\u4e3b\u8981\u4ecb\u7ecd\u4e00\u4e0b\u7b2c2\u79cd\u65b9\u6cd5\uff0c\u7aef\u5230\u7aef\u7684\u8fd0\u52a8\u9884\u6d4b\u3002\u8fd9\u4efb\u52a1\u4ee5\u591a\u65f6\u95f4\u7684\u4f20\u611f\u5668\u6570\u636e\u4f5c\u4e3a\u8f93\u5165\uff0c\u8f93\u51fa\u672a\u6765\u65f6\u523b\u7684\u8fd0\u52a8\u9884\u6d4b\uff08\u5305\u62ec\u7269\u4f53\u68c0\u6d4b\u3001\u8bed\u4e49\u5206\u5272\u3001instance flow\u7b49\uff09\u3002\u5176\u4e2dinstance flow\u7684\u6548\u679c\u56fe\u5927\u6982\u957f\u8fd9\u6837\uff1a<\/p>\n

<\/p>\n

2. \u8054\u5408\u611f\u77e5\u9884\u6d4b\u7684\u5de5\u4f5c\u6d41\u7a0b<\/p>\n

\u603b\u7684\u6765\u8bf4\uff0c\u4e5f\u53ef\u4ee5\u5206\u4e3aencoder-decoder\u6d41\u7a0b\uff1a<\/p>\n

<\/p>\n

\uff081\uff09\u5148\u7528\u4e00\u4e2aencoder\u505a\u6570\u636e\u9884\u5904\u7406\uff0c\u878d\u5408\u591a\u89c6\u89d2\u3001\u591a\u65f6\u95f4\u4fe1\u606f\u4ee5\u5f62\u6210\u4e00\u4e2a\u7efc\u5408\u7684\u7279\u5f81\u8868\u793a\u3002\u7136\u540e\uff082\uff09\u7528\u4e00\u4e2adecoder\uff0c\u7528\u4e00\u4e2a\u9884\u6d4b\u673a\u5236\u9884\u6d4b\u672a\u6765\u7684\u7279\u5f81\u8868\u793a\uff0c\u518d\u89e3\u7801\u6210\u672a\u6765\u7684\u8fd0\u52a8\u9884\u6d4b\u3002\u4e0b\u9762\u5206\u522b\u9610\u8ff0\u5404\u4e2a\u4e3b\u8981\u6210\u5206\uff08\u9ed1\u4f53\u90e8\u5206\uff09<\/p>\n

2.1 \u6570\u636e\u9884\u5904\u7406<\/p>\n

\u5148\u6765\u8bf4\u8bf4\u4f20\u611f\u5668\u6570\u636e\u662f\u56fe\u50cf\u7684\u60c5\u51b5\uff0c\u7136\u540e\u96f7\u8fbe\u3001\u591a\u4f20\u611f\u5668\u878d\u5408\u7684\u60c5\u51b5\u4e5f\u5c31\u5f88\u7c7b\u4f3c\u4e86\u3002<\/p>\n

\u4e00\u822c\u6c7d\u8f66\u7684\u56fe\u50cf\u611f\u77e5\u6570\u636e\u662f\u591a\u89c6\u89d2\uff086\u4e2a\u76f8\u673a\uff0c\u611f\u77e5360\u00b0\u5168\u666f\u4fe1\u606f\uff09\u3001\u65f6\u95f4\u5e8f\u5217\u7684\uff0c I_t = {X^1_ t , …, X^n_ t }, t=1,…,T<\/p>\n

<\/p>\n

\u6240\u4ee5\u9700\u8981\u5148\u878d\u5408\u5404\u4e2a\u89c6\u89d2\u7684\u4fe1\u606f\uff0c\u518d\u878d\u5408\u4e0d\u540c\u65f6\u95f4\u4e0b\u7684\u4fe1\u606f\uff0c\u65b9\u4fbf\u540e\u9762\u505a\u9884\u6d4b\u3002<\/p>\n

\u5bf9\u4e8e\u878d\u5408\u591a\u89c6\u89d2\uff0c\u5e38\u7528\u65b9\u6cd5\u6709\uff08\u5177\u4f53\u5185\u5bb9\u53ef\u4ee5\u53c2\u8003\u539f\u8bba\u6587\uff0c\u6211\u8fd9\u91cc\u5c31\u7b80\u5355\u63cf\u8ff0\u4e00\u4e0b\uff09\uff1a<\/p>\n

\uff081\uff09LSS [19]<\/p>\n

\u5c06\u5404\u4e2a\u89c6\u89d2\u4e0b\u7684\u4e8c\u7ef4\u56fe\u50cf\u62ac\u5347\u50123D\u70b9\u4e91\uff0c\u518d\u538b\u7f29\u6210\u4e8c\u7ef4\u9e1f\u77b0\u56fe\uff08Bird’s eye view\uff0cBEV\uff09\u3002<\/p>\n

\uff082\uff09\u6ce8\u610f\u529b\u673a\u5236 [20]<\/p>\n

\u4e00\u822c\u6cbf\u7528BEVFormer\u5f00\u521b\u7684spatial cross-attention\u65b9\u6cd5\uff1a\u5148\u7528CNN\u6216\u8005Transformer encode\u63d0\u524d\u56fe\u50cf\u6570\u636e\u7279\u5f81\uff0c\u518d\u8bad\u7ec3\u4e00\u4e2aquery\uff0c\u7528\u6ce8\u610f\u529b\u673a\u5236\u63d0\u53d6\u56fe\u50cf\u7279\u5f81\u4fe1\u606f\u3002<\/p>\n

\u4e00\u822c\u8fd9\u4e2a\u8fc7\u7a0b\u4f1a\u5faa\u73af\u505a\u51e0\u6b21\uff0c\u5373\u62ff\u524d\u4e00\u6b21\u8f93\u51fa\u7684bev feature map\u505aquery\uff0c\u518d\u63d0\u53d6\u56fe\u50cf\u7279\u5f81\u4fe1\u606f\uff0c\u518d\u8f93\u51fa\u66f4\u65b0\u7684bev feature map<\/p>\n

\u5bf9\u6bcf\u4e00\u4e2a\u65f6\u523b\u7684\u591a\u89c6\u89d2\u56fe\uff0c\u90fd\u8fd9\u6837\u5904\u7406\uff0c\u6211\u4eec\u5c31\u80fd\u5f97\u5230\u6570\u636e\u7684\u65f6\u5e8f\u7279\u5f81\u8868\u793a S = [B^{(0)}, B^{(\u22121)}, …, B^{(\u2212T )}] \u2208 R^{(T +1)\u00d7X\u00d7Y \u00d7C}<\/p>\n

\u5bf9\u4e8e\u96f7\u8fbe\u6570\u636e\u3001\u591a\u4f20\u611f\u5668\u6570\u636e [1,4,8,11,13,14]\uff0c\u53ea\u662f\u8fd9\u4e00\u6b65\u6709\u533a\u522b\u3002\u7528\u4e0d\u540c\u7684\u65b9\u6cd5\u9884\u5904\u7406\u5b8c\u4e0d\u540c\u6570\u636e\u3001\u878d\u5408\u5b8c\u4e0d\u540c\u7684\u6570\u636e\u4e4b\u540e\uff0c\u4e5f\u662f\u5f97\u5230\u4e00\u4e2a\u65f6\u5e8f\u7279\u5f81\u8868\u793a S = [B^{(0)}, B^{(\u22121)}, …, B^{(\u2212T )}] \uff0c\u7136\u540e\u4e0b\u9762\u7684\u6d41\u7a0b\u5c31\u5b8c\u5168\u4e00\u6837\u4e86\u3002<\/p>\n

2.2 \u878d\u5408\u591a\u89c6\u89d2\u3001\u591a\u65f6\u95f4\u4fe1\u606f<\/p>\n

\u4f46\u4e0a\u9762\u53ea\u662f\u7b80\u5355\u878d\u5408\u4e86\u4e00\u4e0b\u591a\u89c6\u89d2\u4fe1\u606f\uff0c\u5e76\u4e14\u628a\u65f6\u5e8f\u7279\u5f81\u62fc\u63a5\u8d77\u6765\uff0c\u6240\u4ee5\u8fd8\u9700\u8981\u8fdb\u4e00\u6b65\u5904\u7406\u3002\u5e38\u7528\u65b9\u6cd5\u6709\uff1a<\/p>\n

\uff081\uff093D \u5377\u79ef [1,2,9]<\/p>\n

\u56e0\u4e3a\u65f6\u5e8f\u7279\u5f81\u8868\u793a\u662f\u4e2a4\u7ef4\u6570\u636e\uff0c\u6240\u4ee5\u53ef\u4ee5\u75283D\u5377\u79ef\u3002<\/p>\n

\uff082\uff092D \u5377\u79ef \uff08\u4ec5\u9650\u96f7\u8fbe\u6570\u636e\uff09[8]<\/p>\n

3D \u5377\u79ef\u8ba1\u7b97\u91cf\u6709\u70b9\u5927\uff0c\u6240\u4ee5\u53ef\u4ee5\u628a\u65f6\u95f4\u7ef4\u5ea6\u4e5f\u5f53\u6210\u7279\u5f81\u7ef4\u5ea6\uff0c\u8fd9\u6837\u5f97\u5230\u7684\u65f6\u5e8f\u7279\u5f81\u8868\u793a\u5c31\u662f\u4e2a3\u7ef4\u7684\u4e86 S = [B^{(0)}, B^{(\u22121)}, …, B^{(\u2212T )}] \u2208 R^{(C(T +1))\u00d7X\u00d7Y} \uff0c\u5c31\u53ef\u4ee5\u7528\u6b63\u5e38\u7684CNN\u4e86\u3002<\/p>\n

\uff083\uff09\u81ea\u6ce8\u610f\u529b\u673a\u5236 [3]<\/p>\n

\u8fd9\u91cc\u53ef\u4ee5\u50cfBEVFormer\u4e00\u6837\u7528temporal cross-attention\uff0c\u4e5f\u53ef\u4ee5\u7528MotionNet\u3001TBP-Former\u4e2d\u7684Spatial-Temporal Pyramid Network\uff1a<\/p>\n

<\/p>\n

\u5bf9\u6bcf\u4e2a\u901a\u9053\u6570\u636e S_c = [B^{(0)}, B^{(\u22121)}, …, B^{(\u2212T )}]_c \u2208 R^{(T +1)\u00d7X\u00d7Y } \uff0c\u53ef\u4ee5\u5355\u72ec\u53d6\u51fa\u6765\u4f5c\u4e3a\u7279\u5f81\u56fe\uff0c\u7136\u540e\u7528\u81ea\u6ce8\u610f\u529b\u673a\u5236\u5c31\u80fd\u5904\u7406\u3002<\/p>\n

\u5904\u7406\u5b8c\u4e4b\u540e\uff0c\u6211\u4eec\u5c31\u80fd\u5f97\u5230\u4e00\u4e2a\u7efc\u5408\u7684\u7279\u5f81\u8868\u793a\uff0c\u53ef\u4ee5\u662f\u4e00\u4e2a\u7279\u5f81\u56fe\u8868\u793a\u6240\u6709\u8fc7\u53bb\u65f6\u95f4\u7684\u56fe\u50cf S=s in R^{H” times W” times C_{out}} \uff08\u5f53\u6211\u4eec\u7528\uff081\uff09\u6216\uff082\uff09\u7684\u65b9\u6cd5\u5904\u7406\u65f6\uff09\uff0c\u4e5f\u53ef\u4ee5\u662f\u6bcf\u4e00\u4e2a\u65f6\u523b\u90fd\u6709\u4e00\u4e2a\u7279\u5f81\u8868\u793a S \u2208 R^{(T +1)\u00d7X\u00d7Y \u00d7C} \uff08\u5f53\u6211\u4eec\u7528\uff083\uff09\u7684\u65b9\u6cd5\u5904\u7406\u65f6\uff09\u3002<\/p>\n

\u5176\u5b9e\u8fd9\u4e00\u6b65\u505a\u5b8c\u4e4b\u540e\uff0c\u5c31\u53ef\u4ee5\u76f4\u63a5\u505a\u611f\u77e5\u7c7b\u4efb\u52a1\u4e86\uff08\u6bd4\u5982\u5f53\u524d\u65f6\u523b\u7684\u8bed\u4e49\u5206\u5272\u3001\u68c0\u6d4b\u7b49\uff09\uff0c\u5728\u7279\u5f81\u8868\u793a\u540e\u9762\u52a0\u4e0a\u51e0\u4e2a\u4efb\u52a1\u76f8\u5173\u7684decoder head\u5c31\u884c\uff0c\u4f46\u5982\u679c\u6211\u4eec\u8981\u505a\u9884\u6d4b\u7c7b\u4efb\u52a1\uff0c\u90a3\u4e48\u8fd8\u8981\u518d\u9884\u6d4b\u672a\u6765\u65f6\u523b\u7684\u7279\u5f81\u8868\u793a\u3002<\/p>\n

2.3 \u9884\u6d4b\u672a\u6765\u7684\u7279\u5f81\u8868\u793a<\/p>\n

\u63a5\u4e0b\u6765\u5c31\u8981\u9884\u6d4b\u672a\u6765\u65f6\u523b\u7684\u65f6\u5e8f\u7279\u5f81\u8868\u793a\u4e86\u3002\u5e38\u7528\u65b9\u6cd5\u6709\uff1a<\/p>\n

\uff081\uff092D \u5377\u79ef [1]<\/p>\n

\u76f4\u63a5\u63a5\u4e00\u4e2a\u8fc7\u53bb\u65f6\u95f4\u7684\u7279\u5f81\u8868\u793a\uff0c\u8bad\u7ec3\u4e00\u4e2aCNN\uff0c\u8f93\u51fa\u672a\u6765\u65f6\u523b\u7684\u65f6\u5e8f\u7279\u5f81\u8868\u793a\uff0c\u7136\u540e\u7aef\u5230\u7aef\u8bad\u7ec3\u3002\u4f46\u8fd9\u79cd\u65b9\u6cd5\u6548\u679c\u7565\u5dee\u3002<\/p>\n

<\/p>\n

\uff082\uff09RNN\u7c7b\u65b9\u6cd5 [2,4,9]<\/p>\n

RNN\u7c7b\u65b9\u6cd5\u662f\u5f88\u9002\u5408\u505a\u65f6\u5e8f\u4efb\u52a1\u4e86\u3002\u8f93\u5165\u4e00\u4e2a\u8fc7\u53bb\u65f6\u95f4\u7684\u7279\u5f81\u8868\u793a\uff0c\u4f9d\u6b21\u8f93\u51fa\u6bcf\u4e2a\u672a\u6765\u65f6\u523b\u7684\u65f6\u5e8f\u7279\u5f81\u8868\u793a\u3002\u5f88\u591a\u8bba\u6587\u80fd\u628a\u8fd9\u5757\u5b8c\u51fa\u82b1\u6765\uff0c\u52a0\u4e86\u5f88\u591a\u590d\u6742\u7684\u6280\u5de7\uff0c\u4f46\u6211\u8fd9\u91cc\u5c31\u4e0d\u8fc7\u591a\u4ecb\u7ecd\u4e86\u3002<\/p>\n

<\/p>\n

\u4e00\u822c\u8fd9\u79cd\u65b9\u6cd5\u4e5f\u5f88\u559c\u6b22\u7ed3\u5408\u4e00\u4e0b\u6982\u7387\u9884\u6d4b\u673a\u5236<\/p>\n

\uff083\uff09\u6ce8\u610f\u529b\u673a\u5236 [3]<\/p>\n

\u8bad\u7ec3\u4e00\u4e2aquery\uff0c\u901a\u8fc7\u6ce8\u610f\u529b\u673a\u5236\uff0c\u4ece\u8fc7\u53bb\u7684\u65f6\u5e8f\u7279\u5f81\u56fe\u4e2d\u63d0\u53d6\u4fe1\u606f\uff1a<\/p>\n

\u987a\u4fbf\u63d0\u4e00\u4e0b\u4e5f\u6709\u7528\u6982\u7387\u9884\u6d4b\u673a\u5236\u505a\u9884\u6d4b\u7684 [2,6,7,10]\uff0c\u4e5f\u5f88\u6709\u610f\u601d\u3002<\/p>\n

\u56e0\u4e3a\u672a\u6765\u7269\u4f53\u7684\u8fd0\u52a8\u53ef\u80fd\u662f\u4e0d\u786e\u5b9a\u7684\uff0c\u6709\u4e00\u5b9a\u7684\u968f\u673a\u6027\uff0c\u6240\u4ee5\u7528\u6982\u7387\u65b9\u6cd5\u5efa\u6a21\u4e5f\u6709\u5408\u7406\u6027\u3002\u795e\u5947\u7684\u662f\uff0c\u7ecf\u8fc7\u4e00\u7cfb\u5217\u590d\u6742\u7684\u6570\u5b66\u63a8\u5bfc\u4ee5\u53ca\u505a\u4e86\u4e00\u4e9b\u5047\u8bbe\u4e4b\u540e\uff0c\u6982\u7387\u7684\u65b9\u6cd5\u5f80\u5f80\u80fd\u7b80\u5316\u6210 \u5728\u539f\u6765\u7684\u9884\u6d4b\u673a\u5236\u4e2d\u989d\u5916\u52a0\u4e00\u4e2a\u8f93\u5165\uff0c\u8fd9\u4e2a\u8f93\u5165\u662f\u4ece\u4e00\u4e2a\u201c\u672a\u6765\u6982\u7387\u5206\u5e03\uff08\u6a21\u62df\u4e86\u672a\u6765\u5404\u79cd\u53ef\u80fd\u7684\u6982\u7387\u5206\u5e03\uff09\u201d\u4e2d\u62bd\u7684\u4e00\u4e2a\u6837\u672c\uff08\u5305\u542b\u4e86\u672a\u6765\u4fe1\u606f\uff09\u3002<\/p>\n

\u6570\u5b66\u63a8\u5bfc\u6709\u70b9\u590d\u6742\uff0c\u5185\u5bb9\u6bd4\u8f83\u591a\uff0c\u4f46\u5f88\u7cbe\u5f69\uff0c\u6211\u6253\u7b97\u53e6\u5916\u5355\u72ec\u5199\u4e00\u7bc7\u6587\u7ae0\u4ecb\u7ecd\uff01\uff08\u672a\u5b8c\u5f85\u7eed\uff09<\/p>\n

2.4 \u89e3\u7801\u6210\u5404\u79cd\u4efb\u52a1<\/p>\n

\u6709\u4e86\u672a\u6765\u7684\u7279\u5f81\u8868\u793a\uff0c\u518d\u90a3\u4e4b\u540e\u63a5\u4e00\u4e9bdecoder heads\u5c31\u884c\uff08\u6bd4\u5982detection heads\uff09\uff0c\u7136\u540e\u5c31\u80fd\u53bb\u505a\u5404\u79cd\u9884\u6d4b\u4efb\u52a1\u4e86\u3002<\/p>\n

\u5e38\u89c1\u7684\u9884\u6d4b\u4efb\u52a1\u6709\uff1a<\/p>\n

semantic segmentation, instance centerness and instance offset, and future instance flow ([2])<\/p>\n

… + HD map prediction ([3])<\/p>\n

object detection (for generating tracklets from sequential detection) ([1])<\/p>\n

future sementic segmentation, depth and optical flow ([9])<\/p>\n

2.5 \u6a21\u578b\u8bc4\u4f30<\/p>\n

\u6570\u636e\u96c6<\/p>\n

\u6ca1\u6709\u73b0\u6210\u7684\u6570\u636e\u96c6\u6765\u8bc4\u4f30\uff0c\u6240\u4ee5\u4e00\u822c\u662f\u628a\u5176\u5b83\u6570\u636e\u96c6\u6539\u9020\u4e00\u4e0b\u6765\u7528\u3002\u6bd4\u5982\u81ea\u52a8\u9a7e\u9a76\u4e2d\u5e38\u7528\u7684NuScenes [21]\u3001Lyft [22] \u6570\u636e\u96c6\uff0c\u5b83\u4eec\u90fd\u662f\u7ed9\u5355\u4e2a\u6a21\u5757\u8bc4\u4f30\u7528\u7684\uff0c\u6bd4\u5982\u611f\u77e5\u6a21\u5757\u662f\u7528\u5f53\u524d\u65f6\u523b\u7684\u56fe\u50cf\u6765\u505a\u5f53\u524d\u65f6\u523b\u7684\u8bed\u4e49\u5206\u5272\uff0c\u90a3\u4e48\u6211\u4eec\u5c31\u53ef\u4ee5\u7528\u8fc7\u53bb\u65f6\u523b\u7684\u56fe\u50cf\u505a\u8f93\u5165\uff0c\u7136\u540e\u5bf9\u672a\u6765\u65f6\u523b\u7684\u56fe\u50cf\u505a\u8bed\u4e49\u5206\u5272\u3002<\/p>\n

\u8bc4\u4f30\u6807\u51c6\uff08evaluation metric\uff09<\/p>\n

\u5e38\u7528\u7684\u6709\uff081\uff09Intersection over Union (IoU)\uff1a\u672a\u6765\u6bcf\u4e00\u5e27\u56fe\u50cf\u7684\u8bed\u4e49\u5206\u5272\uff0c\uff082\uff09Video Panoptic Quality (VPQ)\uff1a\u5ea6\u91cf\u672a\u6765\u5404\u4e2a\u5e27\u7684\u7269\u4f53\u68c0\u6d4b\u7ed3\u679c\u7684\u8fde\u7eed\u6027<\/p>\n

<\/p>\n

\u8bc4\u4f30\u7ed3\u679c<\/p>\n

\u76ee\u524d\u7684SOTA\u4f3c\u4e4e\u662fTBP-Former [3]\uff0c\u4e00\u4e2a\u7528\u6ce8\u610f\u529b\u673a\u5236\u505a\u7279\u5f81\u878d\u5408\u548c\u9884\u6d4b\u7684\u6a21\u578b\u3002<\/p>\n

<\/p>\n

Reference<\/p>\n

[1] Fast and Furious: Real Time End-to-End 3D Detection, Tracking and Motion Forecasting with a Single Convolutional Net.<\/p>\n

[2] FIERY: Future Instance Prediction in Bird’s-Eye View from Surround Monocular Cameras.<\/p>\n

[3] TBP-Former: Learning Temporal Bird’s-Eye-View Pyramid for Joint Perception and Prediction in Vision-Centric Autonomous Driving.<\/p>\n

[4] PnPNet: End-to-End Perception and Prediction with Tracking in the Loop.<\/p>\n

[5] FISHING Net: Future Inference of Semantic Heatmaps In Grids.<\/p>\n

[6] BEVerse: Unified Perception and Prediction in Birds-Eye-View for Vision-Centric Autonomous Driving.<\/p>\n

[7] StretchBEV: Stretching Future Instance Prediction Spatially and Temporally.<\/p>\n

[8] IntentNet: Learning to Predict Intention from Raw Sensor Data.<\/p>\n

[9] ST-P3: End-to-end Vision-based Autonomous Driving via Spatial-Temporal Feature Learning.<\/p>\n

[10] Implicit Latent Variable Model for Scene-Consistent Motion Forecasting.<\/p>\n

[11] Perceive, Predict, and Plan: Safe Motion Planning Through Interpretable Semantic Representations.<\/p>\n

[12] Learning Interpretable End-to-End Vision-Based Motion Planning for Autonomous Driving with Optical Flow Distillation.<\/p>\n

[13] Deep Multi-Task Learning for Joint Localization, Perception, and Prediction.<\/p>\n

[14] MotionNet: Joint Perception and Motion Prediction for Autonomous Driving Based on Bird’s Eye View Maps.<\/p>\n

[15] Perceive, Interact, Predict: Learning Dynamic and Static Clues for End-to-End Motion Prediction.<\/p>\n

[16] VAD: Vectorized Scene Representation for Efficient Autonomous Driving.<\/p>\n

[17] Computer Vision for Autonomous Vehicles: Problems, Datasets and State of the Art.<\/p>\n

[18] A Survey of End-to-End Driving: Architectures and Training Methods.<\/p>\n

[19] Lift, Splat, Shoot: Encoding Images from Arbitrary Camera Rigs by Implicitly Unprojecting to 3D.<\/p>\n

[20] BEVFormer: Learning Bird’s-Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers.<\/p>\n

[21] nuScenes: A multimodal dataset for autonomous driving.<\/p>\n

[22] One Thousand and One Hours: Self-driving Motion Prediction Dataset.<\/p>\n","protected":false},"excerpt":{"rendered":"

1. \u4ec0\u4e48\u662f\u8054\u5408\u611f\u77e5\u9884\u6d4b\uff1f 1.1 \u81ea\u52a8\u9a7e\u9a76\u7684 […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3,1],"tags":[],"class_list":["post-5565","post","type-post","status-publish","format-standard","hentry","category-technology-frontier","category-home"],"views":15,"_links":{"self":[{"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts\/5565"}],"collection":[{"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/comments?post=5565"}],"version-history":[{"count":2,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts\/5565\/revisions"}],"predecessor-version":[{"id":5567,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts\/5565\/revisions\/5567"}],"wp:attachment":[{"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/media?parent=5565"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/categories?post=5565"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/tags?post=5565"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}