{"id":28140,"date":"2025-04-24T17:43:43","date_gmt":"2025-04-24T09:43:43","guid":{"rendered":"http:\/\/192.168.10.115\/?p=28140"},"modified":"2025-04-24T17:43:43","modified_gmt":"2025-04-24T09:43:43","slug":"2025-04-24-%e9%97%ae%e9%bc%8e%e9%97%ad%e7%8e%afsota%ef%bc%81hip-ad%ef%bc%9a%e5%88%86%e5%b1%82%e5%a4%9a%e7%b2%92%e5%ba%a6%e8%a7%84%e5%88%92%e5%8a%a9%e5%8a%9b%e7%ab%af%e5%88%b0%e7%ab%af%e6%9a%b4","status":"publish","type":"post","link":"http:\/\/222.128.65.89:18086\/index.php\/2025\/04\/24\/28140\/","title":{"rendered":"2025-04-24 \u95ee\u9f0e\u95ed\u73afSOTA\uff01HiP-AD\uff1a\u5206\u5c42\u591a\u7c92\u5ea6\u89c4\u5212\u52a9\u529b\u7aef\u5230\u7aef\u66b4\u6da820\u4e2a\u70b9~"},"content":{"rendered":"\n

\u539f\u6587\u94fe\u63a5\uff1ahttps:\/\/mp.weixin.qq.com\/s\/mqDKtk40i8q9Dv06fTFGpg<\/a><\/p>\n\n\n\n

\u300aHiP-AD: Hierarchical and Multi-Granularity Planning with Deformable Attention for Autonomous Driving in a Single Decoder\u300b\u662f\u4e00\u7bc7\u53d1\u8868\u4e8e 2025 \u5e74 3 \u6708\u7684\u8bba\u6587\uff0c\u6765\u81ea\u4e8enullmax.ai, \u65e8\u5728\u63d0\u5347\u7aef\u5230\u7aef\u81ea\u52a8\u9a7e\u9a76\u7cfb\u7edf\u5728\u95ed\u73af\u8bc4\u4f30\u4e2d\u7684\u6027\u80fd\uff0c\u7279\u522b\u662f\u5728\u89c4\u5212\u6a21\u5757\u7684\u67e5\u8be2\u8bbe\u8ba1\u4e0e\u4ea4\u4e92\u65b9\u9762~<\/p>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2503.08612<\/p>\n\n\n\n

<\/strong>\u80cc\u666f<\/strong><\/strong><\/h2>\n\n\n\n

\u9996\u5148\uff0c\u5927\u591a\u6570\u65b9\u6cd5 UniAD, DriveTransformer, VAD, Para-drive, SparseAD \u5c06\u7aef\u5230\u7aef\u81ea\u52a8\u9a7e\u9a76\u4efb\u52a1\u5f62\u5f0f\u5316\u4e3a\u4e00\u79cd\u901a\u8fc7\u8f68\u8ff9\u56de\u5f52\u8fdb\u884c\u7684\u6a21\u4eff\u5b66\u4e60\u4efb\u52a1<\/strong>\uff0c\u4e14\u76d1\u7763\u4fe1\u53f7\u8f83\u4e3a\u7a00\u758f\uff0c\u4e3b\u8981\u5173\u6ce8\u8f68\u8ff9\u7684\u62df\u5408\uff0c\u800c\u975e\u95ed\u73af\u63a7\u5236\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u9762\u5411\u95ed\u73af\u7684\u81ea\u52a8\u9a7e\u9a76\u65b9\u6cd5\u5219\u9762\u4e34\u5176\u4ed6\u6311\u6218\uff0c\u4f8b\u5982\u975e\u51f8\u4f18\u5316\u95ee\u9898<\/strong> \u548c\u8f6c\u5411\u8bef\u5dee<\/strong>\u3002<\/p>\n\n\n\n

CarLLaVA \u901a\u8fc7\u5c06\u6807\u51c6\u822a\u70b9\u89e3\u8026\u4e3a\u65f6\u95f4\u7ea6\u675f<\/strong>\u548c\u7a7a\u95f4\u7ea6\u675f<\/strong>\u7684\u822a\u70b9\uff0c\u7528\u4e8e\u7eb5\u5411\u548c\u6a2a\u5411\u63a7\u5236\uff0c\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u7f13\u89e3\u4e86\u4e0a\u8ff0\u95ee\u9898\u3002\u7136\u800c\uff0cCarLLaVA \u6784\u5efa\u5728\u4e00\u4e2a\u9884\u8bad\u7ec3\u7684\u5927\u8bed\u8a00\u6a21\u578b\u4e4b\u4e0a\uff0c\u7f3a\u4e4f\u4e2d\u95f4\u7684\u611f\u77e5\u7ed3\u679c\uff0c\u56e0\u6b64\u53ef\u89e3\u91ca\u6027\u4e0d\u8db3\uff0c\u5e76\u4e14\u672a\u63a2\u8ba8\u8f68\u8ff9\u591a\u6837\u6027\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n

\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd \u591a\u7c92\u5ea6\u7684\u89c4\u5212\u67e5\u8be2\u8868\u793a\u65b9\u5f0f\uff0c\u7ed3\u5408\u5206\u5c42\u822a\u70b9\u9884\u6d4b\u673a\u5236\u7528\u4e8e\u7aef\u5230\u7aef\u81ea\u52a8\u9a7e\u9a76\uff0c\u5982\u56fe(c-d)\u6240\u793a\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6211\u4eec\u5c06\u822a\u70b9\u62c6\u5206\u4e3a\uff1a\u65f6\u95f4\u7ef4\u5ea6\u3001\u7a7a\u95f4\u8def\u5f84\u7ef4\u5ea6\u3001\u4ee5\u53ca\u9a7e\u9a76\u98ce\u683c\u7ef4\u5ea6\u7684\u9884\u6d4b\uff0c\u5e76\u5206\u522b\u4f7f\u7528\u76f8\u5e94\u7684\u89c4\u5212\u67e5\u8be2\uff08planning queries\uff09\u6765\u5efa\u6a21\u3002<\/p>\n\n\n\n

\u6b64\u5916\u6587\u7ae0\u5c06\u6bcf\u79cd\u7c7b\u578b\u7684\u822a\u70b9\u591a\u6837\u5316\u4e3a\u591a\u4e2a\u7c92\u5ea6\uff0c\u901a\u8fc7\u4e0d\u540c\u7684\u91c7\u6837\u7b56\u7565\uff08\u5982\u9891\u7387\u3001\u8ddd\u79bb\u548c\u901f\u5ea6\uff09\u8fdb\u884c\u5212\u5206\uff0c\u4ece\u800c\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5f15\u5165\u66f4\u4e30\u5bcc\u7684\u76d1\u7763\u4fe1\u606f\u3002\u8fd9\u4e9b\u591a\u7c92\u5ea6\u822a\u70b9\u80fd\u591f\u88ab\u6709\u6548\u5730\u805a\u5408\uff0c\u4ee5\u4fc3\u8fdb\u4e0d\u540c\u7279\u5f81\u4e4b\u95f4\u7684\u4ea4\u4e92\u3002\u56e0\u6b64\uff0c\u7a00\u758f\u822a\u70b9\u63d0\u4f9b\u4e86\u5168\u5c40\u4fe1\u606f\uff0c\u800c\u5bc6\u96c6\u822a\u70b9\u8f93\u51fa\u5219\u66f4\u9002\u5408\u7528\u4e8e\u7ec6\u7c92\u5ea6\u7684\u63a7\u5236\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0c\u8fd9\u79cd\u591a\u7c92\u5ea6\u8bbe\u8ba1\u663e\u8457\u7f13\u89e3\u4e86\u201c\u81ea\u8f66\u72b9\u8c6b\u201d\u95ee\u9898\u2014\u2014\u5373\u5728\u67d0\u4e9b\u573a\u666f\u4e2d\uff0c\u81ea\u8f66\u6301\u7eed\u7b49\u5f85\uff0c\u76f4\u5230\u95ed\u73af\u4eff\u771f\u8d85\u65f6\u4e3a\u6b62\u3002\u8be5\u65b9\u6cd5\u80fd\u591f\u5728\u4e0d\u5f15\u5165\u56e0\u679c\u7ebf\u7d22\u7684\u60c5\u51b5\u4e0b\uff0c\u9f13\u52b1\u8f66\u8f86\u5728\u590d\u6742\u573a\u666f\uff08\u4f8b\u5982\u4ea4\u901a\u6807\u5fd7\u3001\u8d85\u8f66\u7b49\uff09\u4e2d\u7684\u884c\u4e3a\u5b66\u4e60\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u5176\u6b21\uff0c\u50cf UniAD\u548c VAD\u8fd9\u6837\u7684\u5e8f\u5217\u5f0f\u8303\u5f0f\u4ec5\u4ec5\u5728\u53ef\u5b66\u4e60\u7684 ego \u67e5\u8be2\u4e0e\u611f\u77e5 Transformer \u7684\u8f93\u51fa\u4e4b\u95f4\u5efa\u7acb\u4e86\u4ea4\u4e92\u3002\u800c Para-Drive\u7684\u5e76\u884c\u65b9\u6cd5\u5219\u662f\u8ba9 ego \u67e5\u8be2\u53ea\u4e0e BEV \u7279\u5f81\u8fdb\u884c\u4ea4\u4e92\u3002\u8fd9\u4e9b\u65b9\u6cd5\u90fd\u7f3a\u4e4f\u89c4\u5212\u6a21\u5757\u4e0e\u611f\u77e5\u548c\u573a\u666f\u7279\u5f81\uff08\u5982\u56fe\u50cf\u6216 BEV \u7279\u5f81\uff09\u4e4b\u95f4\u7684\u5168\u9762\u4ea4\u4e92\uff0c\u4ece\u800c\u9650\u5236\u4e86\u89c4\u5212\u7684\u6709\u6548\u6027\u3002<\/p>\n\n\n\n

\u76f8\u6bd4\u4e4b\u4e0b\uff0cDriveTransformer\u5141\u8bb8 ego \u67e5\u8be2\u5728\u4e00\u4e2a\u7edf\u4e00\u7684 Transformer \u4e2d\u540c\u65f6\u4e0e\u611f\u77e5\u7279\u5f81\u548c\u56fe\u50cf\u7279\u5f81\u8fdb\u884c\u5168\u9762\u4ea4\u4e92\u3002\u4f46\u8be5\u65b9\u6cd5\u4f9d\u7136\u9762\u4e34\u6311\u6218\uff1aego \u67e5\u8be2\u5728\u6ca1\u6709\u89c4\u5212\u8f68\u8ff9\u7684\u5148\u9a8c\u4fe1\u606f\u60c5\u51b5\u4e0b\uff0c\u901a\u8fc7\u5168\u5c40\u6ce8\u610f\u529b\u4ece\u591a\u89c6\u89d2\u56fe\u50cf\u4e2d\u63d0\u53d6\u6709\u4ef7\u503c\u4fe1\u606f\u662f\u975e\u5e38\u56f0\u96be\u7684\u3002<\/p>\n\n\n\n

\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u672c\u6587\u63d0\u51fa\u4f7f\u7528\u7ed3\u5408\u7269\u7406\u4f4d\u7f6e\u7684\u89c4\u5212\u53ef\u53d8\u5f62\u6ce8\u610f\u529b\uff08planning deformable attention\uff09\uff0c\u4ee5\u52a8\u6001\u5730\u91c7\u6837\u9760\u8fd1\u89c4\u5212\u8f68\u8ff9\u7684\u56fe\u50cf\u7279\u5f81\u3002\u8be5\u673a\u5236\u53ef\u4ee5\u8f7b\u677e\u5730\u6574\u5408\u8fdb\u7edf\u4e00\u6846\u67b6\u4e2d\uff0c\u5e76\u4e0e\u611f\u77e5\u4efb\u52a1\u534f\u540c\u8fd0\u884c\u3002<\/p>\n\n\n\n

\u5177\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u7edf\u4e00\u89e3\u7801\u5668\uff0c\u5c06\u5305\u542b\u6765\u81ea\u68c0\u6d4b\u4efb\u52a1\uff08\u5305\u62ec\u8fd0\u52a8\u9884\u6d4b\uff09\u3001\u5730\u56fe\u7406\u89e3\u4efb\u52a1\u548c\u89c4\u5212\u4efb\u52a1\u7684\u6df7\u5408\u4efb\u52a1\u67e5\u8be2\uff08hybrid task queries\uff09\u4f5c\u4e3a\u8f93\u5165\u3002\u8fd9\u4f7f\u5f97\u89c4\u5212\u548c\u611f\u77e5\u4efb\u52a1\u80fd\u591f\u5728 BEV \u7a7a\u95f4\u4e2d\u76f8\u4e92\u4ea4\u6362\u4fe1\u606f\uff0c\u540c\u65f6\u901a\u8fc7\u822a\u70b9\u7684\u51e0\u4f55\u5148\u9a8c\uff0c\u4f7f\u89c4\u5212\u67e5\u8be2\u53ef\u4ee5\u4e0e\u56fe\u50cf\u7a7a\u95f4\u4ea4\u4e92\u3002<\/p>\n\n\n\n

<\/strong>\u4e3b\u8981\u8d21\u732e<\/strong><\/strong><\/h2>\n\n\n\n
    \n
  1. \u591a\u7c92\u5ea6\u89c4\u5212\u67e5\u8be2\u8868\u793a\uff08multi-granularity planning query representation<\/strong>)\uff1aHiP-AD \u5f15\u5165\u4e86\u4e00\u79cd\u591a\u7c92\u5ea6\u7684\u89c4\u5212\u67e5\u8be2\u8868\u793a\u65b9\u6cd5\uff0c\u6574\u5408\u4e86\u7a7a\u95f4\u3001\u65f6\u95f4\u548c\u9a7e\u9a76\u98ce\u683c\u7b49\u591a\u79cd\u91c7\u6837\u6a21\u5f0f\u4e0b\u7684\u5f02\u6784\u8def\u6807\u70b9\uff08heterogeneous waypoints\uff09\u3002\u8fd9\u79cd\u8868\u793a\u4e3a\u8f68\u8ff9\u9884\u6d4b\u63d0\u4f9b\u4e86\u989d\u5916\u7684\u76d1\u7763\uff0c\u589e\u5f3a\u4e86\u81ea\u8f66\u7684\u7cbe\u786e\u95ed\u73af\u63a7\u5236\u80fd\u529b\u3002<\/li>\n\n\n\n
  2. \u89c4\u5212\u53ef\u53d8\u5f62\u6ce8\u610f\u529b\u673a\u5236<\/strong>\uff1a\u5229\u7528\u89c4\u5212\u8f68\u8ff9\u7684\u51e0\u4f55\u7279\u6027\uff0cHiP-AD \u7ed3\u5408\u53ef\u53d8\u5f62\u6ce8\u610f\u529b\u673a\u5236\uff0c\u5b9e\u73b0\u4e86\u57fa\u4e8e\u7269\u7406\u4f4d\u7f6e\u7684\u56fe\u50cf\u7279\u5f81\u6709\u6548\u68c0\u7d22\u3002\u8fd9\u79cd\u673a\u5236\u5141\u8bb8\u6a21\u578b\u52a8\u6001\u5730\u4ece\u900f\u89c6\u89c6\u56fe\u4e2d\u63d0\u53d6\u76f8\u5173\u7684\u56fe\u50cf\u7279\u5f81\uff0c\u63d0\u9ad8\u4e86\u611f\u77e5\u7684\u51c6\u786e\u6027\u3002<\/li>\n\n\n\n
  3. \u7edf\u4e00\u89e3\u7801\u5668\u67b6\u6784\uff08UniDecoder\uff09<\/strong>\uff1aHiP-AD \u5728\u4e00\u4e2a\u7edf\u4e00\u7684\u89e3\u7801\u5668\u4e2d\u540c\u65f6\u6267\u884c\u611f\u77e5\u3001\u9884\u6d4b\u548c\u89c4\u5212\u4efb\u52a1\u3002\u901a\u8fc7\u5728\u9e1f\u77b0\u56fe\uff08BEV\uff09\u7a7a\u95f4\u4e2d\uff0c\u89c4\u5212\u67e5\u8be2\u4e0e\u611f\u77e5\u67e5\u8be2\u7684\u8fed\u4ee3\u4ea4\u4e92\uff0c\u5b9e\u73b0\u4e86\u5168\u9762\u7684\u4efb\u52a1\u878d\u5408\u3002<\/li>\n<\/ol>\n\n\n\n

    <\/strong>\u5b9e\u73b0\u65b9\u6cd5<\/strong><\/h2>\n\n\n\n
    \"\"<\/figure>\n\n\n\n

    \u603b\u4f53\u6846\u67b6<\/strong><\/h3>\n\n\n\n

    HiP-AD \u7684\u6574\u4f53\u7f51\u7edc\u7ed3\u6784\u5982\u56fe \u6240\u793a\u3002\u5b83\u7531\u4e00\u4e2a\u9aa8\u5e72\u7f51\u7edc\uff08backbone\uff09\u7ec4\u6210\uff0c\u540e\u63a5\u4e00\u4e2a\u7279\u5f81\u91d1\u5b57\u5854\u7f51\u7edc\uff08FPN\uff09\u6a21\u5757\uff0c\u7528\u4e8e\u4ece\u591a\u89c6\u89d2\u56fe\u50cfI\u4e2d\u63d0\u53d6\u591a\u5c3a\u5ea6\u7279\u5f81F\u3002\u968f\u540e\u662f\u4e00\u4e2a\u7edf\u4e00\u89e3\u7801\u5668\uff08unified decoder\uff09\u4ee5\u53ca\u591a\u4e2a\u7279\u5b9a\u4efb\u52a1\u7684\u9884\u6d4b\u5934\u3002<\/p>\n\n\n\n

    \u8fd9\u4e2a\u7edf\u4e00\u89e3\u7801\u5668\u63a5\u6536\u6df7\u5408\u4efb\u52a1\u951a\u70b9\uff08anchors\uff09\u548c\u67e5\u8be2\uff08queries\uff09\u4f5c\u4e3a\u8f93\u5165\uff0c\u8fd9\u4e9b\u8f93\u5165\u662f\u4ee5\u4e0b\u51e0\u79cd\u67e5\u8be2\u7684\u62fc\u63a5\u7ed3\u679c\uff1a<\/p>\n\n\n\n

      \n
    • \u667a\u80fd\u4f53\u67e5\u8be2\uff08agent queries\uff09<\/strong>\u5bf9\u5e94\u4e8e\u76ee\u6807\u68c0\u6d4b\u548c\u8fd0\u52a8\u9884\u6d4b<\/li>\n\n\n\n
    • \u5730\u56fe\u67e5\u8be2\uff08map queries\uff09<\/strong>\u00a0\u5728\u7ebf\u5730\u56fe\u6784\u5efa<\/li>\n\n\n\n
    • \u89c4\u5212\u67e5\u8be2\uff08planning queries\uff09<\/strong>\u548c\u8f68\u8ff9\u9884\u6d4b\u4efb\u52a1<\/li>\n<\/ul>\n\n\n\n

      \u7f51\u7edc\u4e2d\u7684\u68c0\u6d4b\u548c\u8fd0\u52a8\u9884\u6d4b\u5934\u3001\u5730\u56fe\u7406\u89e3\u5934<\/strong>\u4ee5\u53ca\u89c4\u5212\u5934\u5206\u522b\u7528\u4e8e\u5b8c\u6210\u5404\u81ea\u7684\u4efb\u52a1\u3002\u5176\u4e2d\uff0c\u89c4\u5212\u5934\u4f1a\u8f93\u51fa\u7528\u4e8e\u81ea\u52a8\u9a7e\u9a76\u63a7\u5236\u7684\u65f6\u95f4\u3001\u7a7a\u95f4\u4ee5\u53ca\u9a7e\u9a76\u98ce\u683c\u4e09\u7c7b\u822a\u70b9<\/strong>\u3002\u6b64\u5916\uff0c\u66f4\u65b0\u540e\u7684 top kak_aka\u3001kmk_mkm\u3001kpk_pkp \u4e2a\u67e5\u8be2\u4f1a\u88ab\u4fdd\u5b58\u5728\u8bb0\u5fc6\u6a21\u5757<\/strong>\u4e2d\uff0c\u7528\u4e8e\u540e\u7eed\u7684\u65f6\u5e8f\u4ea4\u4e92\u3002<\/p>\n\n\n\n

      Unified Decoder<\/strong><\/h3>\n\n\n\n
      \"\"<\/figure>\n\n\n\n

      \u7edf\u4e00\u89e3\u7801\u5668\uff08unified decoder\uff09\u7531\u4e09\u4e2a\u6a21\u5757\u7ec4\u6210\uff1a<\/p>\n\n\n\n

        \n
      1. \u65f6\u95f4\u4ea4\u4e92\u6a21\u5757\uff08Temporal Interaction Module\uff09<\/strong><\/li>\n\n\n\n
      2. \u8de8\u4efb\u52a1\u4ea4\u4e92\u6a21\u5757\uff08Collaborative Interaction Module\uff09<\/strong><\/li>\n\n\n\n
      3. \u4efb\u52a1\u53ef\u53d8\u5f62\u805a\u5408\u6a21\u5757\uff08Task Deformable Aggregation Module\uff09<\/strong><\/li>\n<\/ol>\n\n\n\n

        \u6bcf\u4e2a\u6a21\u5757\u5206\u522b\u7528\u4e8e\u652f\u6301\u65f6\u95f4\u7ef4\u5ea6\u4e0a\u7684\u4ea4\u4e92\u3001\u4e0d\u540c\u4efb\u52a1\u4e4b\u95f4\u7684\u534f\u540c\u4ea4\u4e92\uff0c\u4ee5\u53ca\u4efb\u52a1\u4e0e\u56fe\u50cf\u4e4b\u95f4\u7684\u7279\u5f81\u4ea4\u4e92\u3002\u5bf9\u4e8e\u6bcf\u4e2a\u5165\u7684\u4efb\u52a1\u67e5\u8be2\uff08task query\uff09\uff0c\u90fd\u5173\u8054\u6709\u4e00\u4e2a\u5bf9\u5e94\u7684\u951a\u70b9\uff08anchor\uff09\uff1a<\/p>\n\n\n\n

          \n
        • Agent query\uff08\u667a\u80fd\u4f53\u67e5\u8be2\uff09<\/strong>\u00a0\u4f7f\u7528\u6846\u578b\u951a\u70b9\uff08box anchors\uff09Aa<\/li>\n\n\n\n
        • Map query\uff08\u5730\u56fe\u67e5\u8be2\uff09<\/strong>\u00a0\u4f7f\u7528\u6298\u7ebf\u578b\u951a\u70b9\uff08polyline anchors\uff09\u8868\u793a\uff0cAm<\/li>\n\n\n\n
        • Planning query\uff08\u89c4\u5212\u67e5\u8be2\uff09<\/strong>\u00a0\u540c\u6837\u5efa\u6a21\u4e3a\u6298\u7ebf\u578b\u951a\u70b9\uff0c\u901a\u8fc7\u5176\u672a\u6765 T\u4e2a\u822a\u70b9\uff08waypoints\uff09\u6784\u5efa\u800c\u6210\u3002Ap<\/li>\n<\/ul>\n\n\n\n

          \u65f6\u95f4\u4ea4\u4e92\uff08Temporal Interaction\uff09<\/strong>\u65f6\u95f4\u4ea4\u4e92\u6a21\u5757\u7528\u4e8e\u5efa\u7acb\u5f53\u524d\u4efb\u52a1\u7279\u5f81\u4e0e\u5386\u53f2\u4efb\u52a1\u7279\u5f81\u4e4b\u95f4\u7684\u901a\u4fe1\u3002\u8fd9\u4e9b\u5386\u53f2\u7279\u5f81\u901a\u8fc7\u524d\u4e00\u5e27\u63a8\u7406\u4e2d top-k \u9009\u62e9\u673a\u5236<\/strong> \u88ab\u4fdd\u7559\u4e0b\u6765\u3002\u5982\u56fe\u5de6\u4e0b\u89d2\u6240\u793a\uff0c\u8be5\u6a21\u5757\u4e3a\u6bcf\u4e2a\u4efb\u52a1\u7684\u65f6\u95f4\u4ea4\u4e92\u8bbe\u8ba1\u4e86\u4e09\u79cd\u4e0d\u540c\u7684 \u8de8\u6ce8\u610f\u529b\u673a\u5236\uff08cross-attention mechanisms\uff09\uff0c\u6b64\u5916\u8fd8\u5305\u542b\u4e00\u4e2a \u989d\u5916\u7684\u8de8\u6ce8\u610f\u529b\u673a\u5236\uff0c\u4e13\u95e8\u7528\u4e8e\u589e\u5f3a \u89c4\u5212\u67e5\u8be2\uff08planning query\uff09 \u4e0e \u5386\u53f2\u611f\u77e5\u67e5\u8be2\uff08temporal perception queries\uff09 \u4e4b\u95f4\u7684\u4ea4\u4e92\uff0c\u91cd\u70b9\u5173\u6ce8\u5386\u53f2\u7684\u5468\u56f4\u73af\u5883\u5143\u7d20\u3002<\/p>\n\n\n\n

          \u534f\u540c\u4ea4\u4e92\uff08Collaborative Interaction\uff09<\/strong>\u534f\u540c\u4ea4\u4e92\u6a21\u5757\u5b9e\u73b0\u4e86\u8de8\u4efb\u52a1\u4ea4\u4e92\u3002\u5b83\u5305\u62ec\u4e09\u4e2a\u72ec\u7acb\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\uff08self-attention mechanisms\uff09\uff0c\u6bcf\u4e2a\u673a\u5236\u4e13\u95e8\u8d1f\u8d23\u4e00\u4e2a\u4efb\u52a1\uff0c\u5e76\u4e14\u8fd8\u5305\u542b\u4e00\u4e2a\u7edf\u4e00\u7684\u81ea\u6ce8\u610f\u529b\u6a21\u5757\uff0c\u7528\u4e8e\u5728\u4e0d\u540c\u4efb\u52a1\u4e4b\u95f4\u8fdb\u884c\u4ea4\u4e92\u3002\u4e0d\u540c\u4e8e\u4f7f\u7528\u5168\u5c40\u6ce8\u610f\u529b\uff08global attention\uff09\uff0c\u6211\u4eec\u4e3a\u6bcf\u5bf9\u67e5\u8be2\u6784\u5efa\u4e86\u4e00\u4e2a\u51e0\u4f55\u6ce8\u610f\u529b\u56fe\uff08geometric attention map\uff09\uff0c\u4ee5\u4fbf\u4e13\u6ce8\u4e8e\u5c40\u90e8\u548c\u76f8\u5bf9\u5143\u7d20\u3002\u4ee5\u611f\u77e5\u67e5\u8be2\uff08perception query\uff09\u4e3a\u4f8b\uff0c\u6211\u4eec\u901a\u8fc7\u7f29\u653e\u8ddd\u79bb\u4f5c\u4e3a\u6ce8\u610f\u529b\u6743\u91cd\uff0c\u52a8\u6001\u8c03\u6574 BEV \u611f\u53d7\u91ce\uff08receptive fields\uff09\u3002<\/p>\n\n\n\n

          \"\"<\/figure>\n\n\n\n

          \u03c4 \u662f\u4e00\u4e2a\u901a\u8fc7 MLP \u51fd\u6570\u4eceQ\u4e2d\u5b66\u4e60\u5f97\u5230\u7684\u53ef\u5b66\u4e60\u7cfb\u6570\u3002D \u8868\u793a\u4e24\u4e2a\u7269\u4f53\u5b9e\u4f8b\u4e4b\u95f4\u7684\u6b27\u51e0\u91cc\u5f97\u8ddd\u79bb\uff0c\u7c7b\u4f3c\u5730\uff0c\u6211\u4eec\u5c06\u6700\u5c0f\u8ddd\u79bb\u7684\u8ba1\u7b97\u6269\u5c55\uff0c\u7528\u4e8e\u751f\u6210\u6ce8\u610f\u529b\u6743\u91cd\uff0c\u901a\u8fc7\u5f15\u5165\u5730\u56fe-\u76ee\u6807\u4f53\uff08map-agent\uff09\u3001\u5730\u56fe-\u5730\u56fe\uff08map-map\uff09\u3001\u76ee\u6807\u4f53-\u5730\u56fe\uff08agent-map\uff09\u951a\u70b9\u4e4b\u95f4\u7684\u4ea4\u4e92\u6765\u5b9e\u73b0\u3002\u5bf9\u4e8e\u89c4\u5212\u67e5\u8be2\uff08planning queries\uff09\uff0c\u6ca1\u6709\u8ddd\u79bb\u9650\u5236\uff0c\u8fd9\u4f7f\u5f97\u5b83\u4eec\u80fd\u591f\u4ece\u6240\u6709\u4efb\u52a1\u4e2d\u83b7\u53d6\u4fe1\u606f\u3002<\/p>\n\n\n\n

          \u4efb\u52a1\u53ef\u53d8\u5f62\u6ce8\u610f\u529b\uff08Task Deformable Attention\uff09<\/strong>\u4f7f\u7528\u5168\u5c40\u6ce8\u610f\u529b\u4e0e\u6240\u6709\u591a\u89c6\u89d2\u56fe\u50cf\u7279\u5f81\u8fdb\u884c\u4ea4\u4e92\u7684\u65b9\u6cd5\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u72ec\u7acb\u7684\u53ef\u53d8\u5f62\u6ce8\u610f\u529b\u6a21\u5757<\/strong>\uff0c\u4e3a\u6bcf\u4e2a\u4efb\u52a1\u67e5\u8be2\u63d0\u53d6\u5c40\u90e8\u7a00\u758f\u7279\u5f81\uff0c\u4ee5\u66f4\u8d34\u5408\u5404\u81ea\u4efb\u52a1\u7684\u9700\u6c42\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u5c06\u4efb\u52a1\u951a\u70b9\uff08task anchors\uff09\u901a\u8fc7\u76f8\u673a\u53c2\u6570\u6295\u5f71\u5230\u591a\u89c6\u89d2\u56fe\u50cf\u4e0a\uff0c\u5bf9\u4e8e\u89c4\u5212\u4efb\u52a1\uff0c\u6211\u4eec\u5c06\u53c2\u8003\u822a\u70b9\uff08reference waypoints\uff09\u5206\u5e03\u5728\u591a\u4e2a\u9884\u5b9a\u4e49\u7684\u9ad8\u5ea6\u503c\u4e0a\uff0c\u5e76\u5c06\u5176\u6295\u5f71\u5230\u591a\u89c6\u89d2\u56fe\u50cf\u4e0a\u3002\u4e3a\u4e86\u91c7\u6837\u5468\u56f4\u70b9\u7684\u7279\u5f81\uff0c\u6211\u4eec\u4f7f\u7528\u591a\u4e2a\u591a\u5c42\u611f\u77e5\u673a\uff08MLP\uff09\u6839\u636e\u6295\u5f71\u540e\u7684\u53c2\u8003\u70b9\u6765\u5b66\u4e60\u7a7a\u95f4\u504f\u79fb\u91cf\uff08spatial offsets\uff09\u548c\u76f8\u5173\u6743\u91cd\uff08associated weights\uff09\u3002\uff08Planning Deformable Attention, PDA\uff09<\/p>\n\n\n\n

          \"\"<\/figure>\n\n\n\n

          P \u662f\u6620\u5c04\u51fd\u6570. \u4ed6\u6574\u5408\u4e86\u672a\u6765\u8f68\u8ff9\u7684\u7279\u5f81\uff0c\u6765\u8868\u8fbe\u573a\u666f\u7279\u5f81\uff0c\u907f\u514d\u78b0\u649e\u98ce\u9669<\/p>\n\n\n\n

          \u5206\u5c42\u4e0e\u591a\u7c92\u5ea6\u89c4\u5212\uff08Hierarchical and Multi-granularity Planning\uff09<\/strong><\/p>\n\n\n\n

          \"\"<\/figure>\n\n\n\n

          \u5206\u5c42\u822a\u70b9\uff08Hierarchical Waypoints\uff09<\/strong>\uff1a\u6587\u7ae0\u4f7f\u7528\u4e86\u65f6\u95f4\uff08temporal\uff09\u548c\u7a7a\u95f4\uff08spatial\uff09\u822a\u70b9\uff0c\u8fd8\u5f15\u5165\u4e86\u9a7e\u9a76\u98ce\u683c\uff08driving-style\uff09\u822a\u70b9\uff0c\u867d\u7136\u9a7e\u9a76\u98ce\u683c\u822a\u70b9\u5728\u5f62\u5f0f\u4e0a\u7c7b\u4f3c\u4e8e\u65f6\u95f4\u822a\u70b9\uff0c\u4f46\u5b83\u4eec\u8fdb\u4e00\u6b65\u878d\u5408\u4e86\u901f\u5ea6\u4fe1\u606f<\/strong>\uff0c\u4ee5\u4fbf\u5728\u590d\u6742\u73af\u5883\u4e2d\u66f4\u597d\u5730\u5b66\u4e60\u81ea\u8f66\u884c\u4e3a\u3002\u6b64\u5916\uff0c\u6587\u7ae0\u8fd8\u5f15\u5165\u4e86\u4e00\u79cd\u591a\u91c7\u6837\u7b56\u7565\uff08multi-sampling strategy\uff09\uff0c\u4ee5\u5b9e\u73b0\u4e30\u5bcc\u7684\u8f68\u8ff9\u76d1\u7763\u548c\u7cbe\u7ec6\u7684\u63a7\u5236\u3002<\/strong>\u8be5\u7b56\u7565\u7ed3\u5408\u4e86\uff1a<\/p>\n\n\n\n

            \n
          • \u7a7a\u95f4\u822a\u70b9\u7684\u7a00\u758f\u4e0e\u5bc6\u96c6\u95f4\u9694\uff1b<\/li>\n\n\n\n
          • \u65f6\u95f4\u4e0e\u9a7e\u9a76\u98ce\u683c\u822a\u70b9\u7684\u9ad8\u9891\u4e0e\u4f4e\u9891\u91c7\u6837\uff1b<\/li>\n\n\n\n
          • \u4e0d\u540c\u901f\u5ea6\u6761\u4ef6\u4e0b\u7684\u9a7e\u9a76\u98ce\u683c\u822a\u70b9\u3002<\/li>\n\n\n\n
          • \u8fd9\u6837\u7684\u8bbe\u8ba1\u4f7f\u5f97\u7a00\u758f\u95f4\u9694\u7684\u822a\u70b9\u53ef\u4ee5\u63d0\u4f9b\u66f4\u5e7f\u6cdb\u7684\u5168\u5c40\u8bed\u4e49\u4e0a\u4e0b\u6587\uff0c\u6709\u52a9\u4e8e\u66f4\u9ad8\u5c42\u6b21\u7684\u51b3\u7b56\u5236\u5b9a\uff1b\u800c\u5bc6\u96c6\u95f4\u9694\u7684\u822a\u70b9\u5219\u652f\u6301\u66f4\u7cbe\u7ec6\u7684\u64cd\u4f5c\u63a7\u5236\u3002<\/li>\n<\/ul>\n\n\n\n

            \u65f6\u95f4\u4e0e\u7a7a\u95f4\u822a\u70b9\u4e4b\u95f4\uff0c\u4ee5\u53ca\u9ad8\u4f4e\u9891\u3001\u7a00\u5bc6\u95f4\u9694\u91c7\u6837\u7b56\u7565\u4e4b\u95f4\uff0c\u662f\u4e92\u8865\u5173\u7cfb<\/strong>\uff0c\u5171\u540c\u6784\u5efa\u4e86\u66f4\u9c81\u68d2\u3001\u66f4\u9ad8\u6548\u7684\u89c4\u5212\u80fd\u529b\u3002\u6b64\u5916\uff0c\u9a7e\u9a76\u98ce\u683c\u822a\u70b9\u7ed3\u5408\u4e0d\u540c\u901f\u5ea6<\/strong>\uff0c\u80fd\u5e2e\u52a9\u6a21\u578b\u66f4\u51c6\u786e\u5730\u7406\u89e3\u5982\u8d85\u8f66\u6216\u7d27\u6025\u5236\u52a8\u7b49\u573a\u666f<\/strong>\uff0c\u5728\u95ed\u73af\u8bc4\u4f30\u4e2d\u63d0\u4f9b\u66f4\u7075\u6d3b\u7684\u7eb5\u5411\u63a7\u5236\u80fd\u529b\u3002<\/p>\n\n\n\n

            \u591a\u7c92\u5ea6\u89c4\u5212\u67e5\u8be2\uff08multi-granularity planning queries\uff09<\/strong>\uff0c\u7528\u4e8e\u9884\u6d4b\u8fd9\u4e9b\u5f02\u6784\u822a\u70b9\uff08heterogeneous waypoints\uff09\u3002 \u65f6\u95f4\u7c92\u5ea6<\/strong>\uff08temporal\uff09\u67e5\u8be2\uff1b\u7a7a\u95f4\u7c92\u5ea6<\/strong>\uff08spatial\uff09\u67e5\u8be2\uff1b\u9a7e\u9a76\u98ce\u683c\u7c92\u5ea6<\/strong>\uff08driving-style\uff09\u67e5\u8be2\u3002\u5f53\u8fd9\u4e9b\u67e5\u8be2\u901a\u8fc7\u7edf\u4e00\u89e3\u7801\u5668\uff08unified decoder\uff09\u5904\u7406\u4e4b\u540e\uff0c\u540c\u4e00\u6a21\u6001\u4e0b\u4e0d\u540c\u7c92\u5ea6\u7684\u89c4\u5212\u67e5\u8be2\u4f1a\u88ab\u5bf9\u9f50\u5e76\u805a\u5408<\/strong>\uff0c\u751f\u6210\u4e00\u4e2a\u878d\u5408\u540e\u7684\u67e5\u8be2\uff08fused query\uff09\uff0c\u4ece\u800c\u589e\u5f3a\u4fe1\u606f\u7684\u4e92\u8865\u6027\u4e0e\u6574\u4f53\u6548\u679c\u3002\u6700\u7ec8\uff0c\u8fd9\u4e2a\u878d\u5408\u67e5\u8be2\u88ab\u7528\u4e8e\u9884\u6d4b\u6240\u6709\u7c92\u5ea6\u4e0b\u7684\u822a\u70b9\uff08waypoints\uff09<\/strong>\uff0c\u5e76\u501f\u52a9\u989d\u5916\u7684\u76d1\u7763\u4fe1\u53f7\u6765\u8fdb\u4e00\u6b65\u4f18\u5316\u8f68\u8ff9\u9884\u6d4b\u3002<\/p>\n\n\n\n

            <\/strong>\u5bf9\u9f50\u5339\u914d\uff08Align Matching\uff09<\/strong><\/strong><\/h2>\n\n\n\n

            \u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u5728\u6bcf\u4e2a\u67e5\u8be2\u7ec4\u5185\u91c7\u7528\u4e86\u8d62\u5bb6\u901a\u5403\uff08winner-takes-all\u7684\u5339\u914d\u65b9\u5f0f\u6765\u9009\u62e9\u6700\u4f18\u6a21\u6001\uff08modality\uff09\u8fdb\u884c\u4f18\u5316\u3002\u6bcf\u4e2a\u67e5\u8be2\u7ec4\u5305\u542b\u5177\u6709\u76f8\u540c\u7c92\u5ea6\uff08granularity\uff09\u7684\u6240\u6709\u6a21\u6001\u3002\u4e0d\u540c\u4e8e\u4e3a\u6bcf\u7ec4\u822a\u70b9\u5355\u72ec\u8fdb\u884c\u5339\u914d\u7684\u505a\u6cd5\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u5bf9\u9f50\u5339\u914d\u7b56\u7565\uff08align matching strategy\uff09<\/strong>\uff0c\u8be5\u7b56\u7565\u6307\u5b9a\u67d0\u4e00\u7ec4\u822a\u70b9\u4f5c\u4e3a\u53c2\u8003\u822a\u70b9<\/strong> \u5e76\u4f7f\u7528\u5176\u5bf9\u5e94\u7684\u771f\u5b9e\u6807\u7b7e<\/strong>\u8fdb\u884c\u5339\u914d\u3002\u6838\u5fc3\u601d\u60f3\u662f\uff1a\u5728\u8bad\u7ec3\u8f68\u8ff9\u9884\u6d4b\u65f6\uff0c\u4e3a\u4e86\u51cf\u5c11\u6a21\u6001\u95f4\u5339\u914d\u6df7\u4e71\u548c\u63d0\u9ad8\u6548\u7387\uff0c\u4e0d\u662f\u5bf9\u6240\u6709\u6a21\u6001\u72ec\u7acb\u5339\u914d\u5e76\u4f18\u5316\uff0c\u800c\u662f\u9009\u5b9a\u4e00\u4e2a\u6700\u4f18\u7684\u6a21\u6001\uff08\u6bd4\u5982\u66f4\u63a5\u8fd1 GT \u7684\u5de6\u8f6c\u3001\u53f3\u8f6c\u6216\u76f4\u884c\u8f68\u8ff9\uff09\u4f5c\u4e3a\u201c\u53c2\u7167\u8f68\u8ff9\u201d\u8fdb\u884c\u8bad\u7ec3\u4f18\u5316\uff0c\u5f3a\u5316\u6a21\u578b\u5bf9\u6b63\u786e\u8f68\u8ff9\u6a21\u5f0f\u7684\u5b66\u4e60\u3002\u8fd9\u6837\u80fd\u5728\u591a\u6a21\u6001\u9884\u6d4b\u4e2d\u805a\u7126\u4e8e\u6700\u6709\u53ef\u80fd\u7684\u771f\u5b9e\u884c\u4e3a\u3002<\/p>\n\n\n\n

            <\/strong>\u5b9e\u9a8c\u7ed3\u679c<\/strong><\/strong><\/h2>\n\n\n\n

            \u5728\u95ed\u73af\u57fa\u51c6\u6d4b\u8bd5 Bench2Drive \u4e0a\uff0cHiP-AD \u8d85\u8d8a\u4e86\u6240\u6709\u73b0\u6709\u7684\u7aef\u5230\u7aef\u81ea\u52a8\u9a7e\u9a76\u65b9\u6cd5\uff0c\u5e76\u5728\u771f\u5b9e\u4e16\u754c\u6570\u636e\u96c6 nuScenes \u4e0a\u53d6\u5f97\u4e86\u5177\u6709\u7ade\u4e89\u529b\u7684\u8868\u73b0\u3002<\/p>\n\n\n\n

            \"\"<\/figure>\n\n\n\n

            \u6280\u672f\u4eae\u70b9<\/strong><\/strong><\/h2>\n\n\n\n
              \n
            • \u89c4\u5212\u67e5\u8be2\u4e0e\u611f\u77e5\u67e5\u8be2\u7684\u4ea4\u4e92<\/strong>\uff1aHiP-AD \u5141\u8bb8\u89c4\u5212\u67e5\u8be2\u4e0e\u611f\u77e5\u67e5\u8be2\u5728 BEV \u7a7a\u95f4\u4e2d\u8fdb\u884c\u8fed\u4ee3\u4ea4\u4e92\uff0c\u589e\u5f3a\u4e86\u7cfb\u7edf\u5bf9\u73af\u5883\u7684\u7406\u89e3\u548c\u9884\u6d4b\u80fd\u529b\u3002<\/li>\n\n\n\n
            • \u52a8\u6001\u7279\u5f81\u63d0\u53d6<\/strong>\uff1a\u901a\u8fc7\u53ef\u53d8\u5f62\u6ce8\u610f\u529b\u673a\u5236\uff0c\u6a21\u578b\u80fd\u591f\u6839\u636e\u89c4\u5212\u8f68\u8ff9\u7684\u7269\u7406\u4f4d\u7f6e\uff0c\u4ece\u900f\u89c6\u89c6\u56fe\u4e2d\u52a8\u6001\u63d0\u53d6\u76f8\u5173\u7684\u56fe\u50cf\u7279\u5f81\uff0c\u63d0\u9ad8\u4e86\u611f\u77e5\u7684\u51c6\u786e\u6027\u3002<\/li>\n\n\n\n
            • \u591a\u4efb\u52a1\u878d\u5408<\/strong>\uff1a\u5728\u4e00\u4e2a\u7edf\u4e00\u7684\u89e3\u7801\u5668\u4e2d\u540c\u65f6\u5904\u7406\u611f\u77e5\u3001\u9884\u6d4b\u548c\u89c4\u5212\u4efb\u52a1\uff0c\u51cf\u5c11\u4e86\u4efb\u52a1\u4e4b\u95f4\u7684\u4fe1\u606f\u635f\u5931\uff0c\u63d0\u9ad8\u4e86\u6574\u4f53\u7cfb\u7edf\u7684\u6548\u7387\u548c\u6027\u80fd\u3002<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"

              \u539f\u6587\u94fe\u63a5\uff1ahttps:\/\/mp.weixin. […]<\/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-28140","post","type-post","status-publish","format-standard","hentry","category-technology-frontier","category-home"],"views":0,"_links":{"self":[{"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts\/28140"}],"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=28140"}],"version-history":[{"count":1,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts\/28140\/revisions"}],"predecessor-version":[{"id":28148,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/posts\/28140\/revisions\/28148"}],"wp:attachment":[{"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/media?parent=28140"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/categories?post=28140"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/222.128.65.89:18086\/index.php\/wp-json\/wp\/v2\/tags?post=28140"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}