{"id":53724,"date":"2025-02-16T08:04:08","date_gmt":"2025-02-16T00:04:08","guid":{"rendered":"https:\/\/fwq.ai\/blog\/53724\/"},"modified":"2025-02-16T08:04:08","modified_gmt":"2025-02-16T00:04:08","slug":"%e5%8a%a0%e9%80%9f%e5%ae%9e%e6%97%b6%e8%a7%86%e8%a7%89%e5%ba%94%e7%94%a8","status":"publish","type":"post","link":"https:\/\/fwq.ai\/blog\/53724\/","title":{"rendered":"\u52a0\u901f\u5b9e\u65f6\u89c6\u89c9\u5e94\u7528"},"content":{"rendered":"<p>\u5728\u8fb9\u7f18\u90e8\u7f72\u673a\u5668\u5b66\u4e60\u901a\u5e38\u4f1a\u5728\u4f20\u611f\u548c\u76d1\u63a7\u5e94\u7528\u4e2d\u63d0\u4f9b\u663e\u8457\u4f18\u52bf\u3002\u901a\u8fc7\u5c06\u5904\u7406\u4fdd\u6301\u5728\u6570\u636e\u6e90\u9644\u8fd1\uff0c\u53ef\u4ee5\u8282\u7701\u5927\u91cf\u7684\u7f51\u7edc\u3001\u5b58\u50a8\u548c\u4e91\u8ba1\u7b97\u6210\u672c\uff0c\u5e76\u4e14\u6574\u4f53\u5ef6\u8fdf\u4e5f\u4f1a\u964d\u4f4e\u3002\u7136\u800c\uff0c\u8fb9\u7f18\u8bbe\u5907\u53d7\u9650\u4e8e\u5176\u6709\u9650\u7684\u8ba1\u7b97\u8d44\u6e90\uff0c\u8fd9\u4e9b\u8d44\u6e90\u901a\u5e38\u6bd4\u670d\u52a1\u5668\u7684\u8981\u5f31\u3002\u56e0\u6b64\uff0c\u5f53\u5728\u8fb9\u7f18\u90e8\u7f72\u673a\u5668\u5b66\u4e60\u5e94\u7528\u7a0b\u5e8f\u65f6\uff0c\u5fc5\u987b\u5bf9\u5176\u8fdb\u884c\u4e13\u95e8\u4f18\u5316\u4ee5\u6700\u5927\u5316\u6027\u80fd\u3002<\/p>\n<p>\u5728\u8fd9\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u5c06\u8ba8\u8bba\u52a0\u901f\u5b9e\u65f6\u8ba1\u7b97\u673a\u89c6\u89c9\u5e94\u7528\u7684\u6280\u672f\uff0c\u8fd9\u4e9b\u6280\u672f\u53ef\u4ee5\u5728\u667a\u80fd\u4f20\u611f\u3001\u76d1\u63a7\u6216\u673a\u5668\u4eba\u7b49\u573a\u666f\u4e2d\u627e\u5230\u3002\u6211\u4eec\u5c06\u901a\u8fc7\u4f7f\u7528\u5f00\u6e90\u5e93\uff08\u5982OpenCV\u3001Roboflow\u63a8\u7406\u548cNvidia Deepstream\uff09\u6765\u5b9e\u73b0\u63a8\u7406\u7ba1\u9053\u3002<\/p>\n<h2>1\u3001\u63a8\u7406\u7ba1\u9053<\/h2>\n<p>\u63a8\u7406\u7ba1\u9053\u5c06\u8fd0\u884c\u89c6\u9891\u6d41\u4e0a\u7684\u673a\u5668\u5b66\u4e60\u63a8\u7406\u7684\u8fc7\u7a0b\u5206\u4e3a\u4e00\u7cfb\u5217\u79bb\u6563\u6b65\u9aa4\u3002\u4e00\u4e2a\u57fa\u672c\u7684\u7ba1\u9053\u901a\u5e38\u5305\u62ec\u4ee5\u4e0b\u9636\u6bb5\uff1a<\/p>\n<ul>\n<li>\u89c6\u9891\u6d41\u6570\u636e\u88ab\u63d0\u53d6\u5e76\u89e3\u7801\u5f62\u6210\u56fe\u50cf\u5e27\u3002<\/li>\n<li>\u5e27\u5728\u63a8\u7406\u524d\u8fdb\u884c\u9884\u5904\u7406\uff08\u8c03\u6574\u5927\u5c0f\u3001\u5f52\u4e00\u5316\u7b49\uff09\u3002<\/li>\n<li>\u6279\u91cf\u5e27\u88ab\u8f6c\u6362\u4e3a\u5f20\u91cf\u5e76\u53d1\u9001\u5230\u8bbe\u5907GPU\uff08\u5982\u679c\u6ca1\u6709\u53ef\u7528GPU\uff0c\u5219\u53d1\u9001\u5230CPU\uff09\u8fdb\u884c\u63a8\u7406\u3002<\/li>\n<li>\u63a8\u7406\u7ed3\u679c\u88ab\u53e0\u52a0\u5230\u539f\u59cb\u5e27\u4e0a\uff0c\u5e76\u663e\u793a\u5728\u5c4f\u5e55\u4e0a\u3001\u672c\u5730\u4fdd\u5b58\u6216\u5411\u4e0b\u4f20\u8f93\u3002<\/li>\n<li>\u63a8\u7406\u7ed3\u679c\u4f5c\u4e3a\u5143\u6570\u636e\u53d1\u9001\u4ee5\u89e6\u53d1\u8fdb\u4e00\u6b65\u64cd\u4f5c\u3002<\/li>\n<\/ul>\n<p>\u66f4\u590d\u6742\u7684\u7ba1\u9053\u53ef\u80fd\u5305\u62ec\u5bf9\u8c61\u8ddf\u8e2a\u3001\u611f\u5174\u8da3\u533a\u57df\u8fc7\u6ee4\u548c\u5176\u4ed6\u4e1a\u52a1\u903b\u8f91\uff0c\u5177\u4f53\u53d6\u51b3\u4e8e\u5e94\u7528\u573a\u666f\u3002<\/p>\n<h2>2\u3001\u63a8\u7406\u7ba1\u9053\u7684\u5f02\u6b65\u5316<\/h2>\n<p>\u867d\u7136\u63a8\u7406\u7ba1\u9053\u4e2d\u7684\u6b65\u9aa4\u662f\u4f9d\u6b21\u6267\u884c\u7684\uff0c\u4f46\u5b83\u4eec\u5b9e\u9645\u4e0a\u53ef\u4ee5\u901a\u8fc7\u4f7f\u7528<strong>\u591a\u7ebf\u7a0b<\/strong>\u5728\u8fd0\u884c\u65f6\u5e76\u53d1\u6267\u884c\u3002<\/p>\n<p>\u591a\u7ebf\u7a0b\u662f\u6307\u5728\u540c\u4e00\u65f6\u95f4\u8fd0\u884c\u591a\u4e2a\u7ebf\u7a0b\u6216\u53ef\u8fd0\u884c\u4ee3\u7801\u6bb5\u7684\u8fc7\u7a0b\u3002\u5728Python\u4e2d\uff0c\u8fd9\u662f\u901a\u8fc7\u4e0a\u4e0b\u6587\u5207\u6362\u5b9e\u73b0\u7684\uff0c\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u5904\u7406\u5668\u5728\u5f53\u524d\u7ebf\u7a0b\u7a7a\u95f2\u65f6\u5207\u6362\u5230\u5176\u4ed6\u7ebf\u7a0b\uff0c\u4ece\u800c\u7ed9\u4eba\u4e00\u79cd\u4e24\u8005\u540c\u65f6\u6267\u884c\u7684\u9519\u89c9\u3002<\/p>\n<p>\u7531\u4e8e\u63a8\u7406\u7ba1\u9053\u5305\u62ecI\/O\u64cd\u4f5c\uff08\u8bfb\u53d6\u5e27\u3001\u5c06\u5f20\u91cf\u590d\u5236\u5230GPU\u3001\u4f20\u8f93\u7ed3\u679c\uff09\uff0c\u5728\u6b64\u671f\u95f4CPU\u5904\u4e8e\u7b49\u5f85\u72b6\u6001\uff0c\u56e0\u6b64\u591a\u7ebf\u7a0b\u63d0\u4f9b\u4e86\u52a0\u901f\u7684\u6f5c\u529b\u3002\u8fd9\u5728\u5e27\u4ee5\u6279\u6b21\u65b9\u5f0f\u5904\u7406\u65f6\u7279\u522b\u6709\u7528\uff0c\u56e0\u4e3a\u8fd9\u53ef\u4ee5\u8ba9CPU\u6709\u65f6\u95f4\u7ee7\u7eed\u5904\u7406\u5176\u4ed6\u4efb\u52a1\uff0c\u800c\u5f53\u6279\u6b21\u6b63\u5728\u6536\u96c6\u548c\u53d1\u9001\u8fdb\u884c\u63a8\u7406\u65f6\u3002<\/p>\n<p>\u8fd9\u5c31\u662fRoboflow\u7684\u63a8\u7406\u7ba1\u9053\u5b9e\u73b0\u7684\u57fa\u7840\uff0c\u5b83\u5c06\u89c6\u9891\u89e3\u7801\u3001\u63a8\u7406\u548c\u540e\u5904\u7406\u5206\u79bb\u5230\u4e13\u7528\u7ebf\u7a0b\u4e2d\u5e76\u53d1\u8fd0\u884c\u3002<\/p>\n<p>  \u591a\u7ebf\u7a0b\u7ba1\u9053\u3002\u6765\u6e90\uff1a <\/p>\n<p>\u8003\u8651\u4ee5\u4e0b\u4f7f\u7528OpenCV\u4e2d\u7684YOLOv8-nano\u76ee\u6807\u68c0\u6d4b\u6a21\u578b\u5b9e\u73b0\u7684\u540c\u6b65\u63a8\u7406\u7ba1\u9053\uff1a<\/p>\n<pre><code>import cv2  \nfrom inference import get_model  \nimport supervision as sv  \n  \nmodel = get_model(\"yolov8n-640\")  \n  \ncap = cv2.VideoCapture('\/path\/to\/video')  \nwhile cap.isOpened():  \n    ret, frame = cap.read()  \n    if not ret:  \n        print(\"Can't receive frame (stream end?). Exiting ...\")  \n        break  \n      \n    predictions = model.infer(frame)  \n  \n    detections = sv.Detections.from_inference(predictions[0])  \n    annotated_image = sv.BoxAnnotator().annotate(scene=frame, detections=detections)  \n    annotated_image = sv.LabelAnnotator().annotate(scene=annotated_image, detections=detections)  \n      \n    cv2.imshow(\"output\", annotated_image)  \n    if cv2.waitKey(1) == ord('q'):  \n        break  \n  \ncap.release()  \ncv2.destroyAllWindows()\n<\/code><\/pre>\n<p>\u5728\u4e0a\u8ff0\u4ee3\u7801\u4e2d\uff0c\u7ba1\u9053\u4ece\u89c6\u9891\u6d41\u4e2d\u8bfb\u53d6\u5e27\uff0c\u8c03\u7528\u68c0\u6d4b\u6a21\u578b\uff0c\u6ce8\u91ca\u5e27\uff0c\u6700\u540e\u5c06\u5176\u6295\u5f71\u5230\u663e\u793a\u3002\u8fd9\u4e9b\u6b65\u9aa4\u662f\u987a\u5e8f\u6267\u884c\u7684\uff0c\u610f\u5473\u7740\u6bcf\u4e2a\u6b65\u9aa4\u5fc5\u987b\u5b8c\u6210\u6267\u884c\u540e\u4e0b\u4e00\u4e2a\u6b65\u9aa4\u624d\u80fd\u5f00\u59cb\u3002<\/p>\n<p>\u4e0e\u4f7f\u7528Roboflow\u63a8\u7406\u7ba1\u9053\u5b9e\u73b0\u76f8\u540c\u7ba1\u9053\u7684\u4ee3\u7801\u76f8\u6bd4\uff1a<\/p>\n<pre><code>from inference import InferencePipeline  \nfrom inference.core.interfaces.camera.entities import VideoFrame  \nimport supervision as sv  \nimport cv2  \n  \ndef on_prediction(  \n    predictions: dict,  \n    video_frame: VideoFrame,  \n) -&gt; None:  \n      \n    detections = sv.Detections.from_inference(predictions)  \n    annotated_image = sv.BoxAnnotator().annotate(scene=frame, detections=detections)  \n    annotated_image = sv.LabelAnnotator().annotate(scene=annotated_image, detections=detections)  \n      \n    cv2.imshow(\"output\", annotated_image)  \n    cv2.waitKey(1)  \n  \n# \u521b\u5efa\u63a8\u7406\u7ba1\u9053\u5bf9\u8c61  \npipeline = InferencePipeline.init(  \n    model_id=\"yolov8n-640\",   \n    video_reference='\/path\/to\/video'  \n)  \n# \u542f\u52a8\u7ba1\u9053  \npipeline.start()  \n# \u7b49\u5f85\u7ba1\u9053\u5b8c\u6210  \npipeline.join()\n<\/code><\/pre>\n<p>\u5728\u4e0a\u9762\u7684\u4f8b\u5b50\u4e2d\uff0c\u5e27\u6ce8\u91ca\u548c\u663e\u793a\u662f\u5728\u4e00\u4e2a\u56de\u8c03\u51fd\u6570\u4e2d\u5b9e\u73b0\u7684\uff0c\u8be5\u51fd\u6570\u5728\u4e00\u6279\u63a8\u7406\u5b8c\u6210\u540e\u6267\u884c\u3002<\/p>\n<p>\u5728Nvidia Jetson AGX Xavier\u4e0a\u8fd0\u884c\u8fd9\u4e24\u4e2a\u7a0b\u5e8f\uff0c\u6211\u4eec\u5f97\u5230\u4ee5\u4e0b\u541e\u5410\u91cf\u6d4b\u91cf\u7ed3\u679c\uff1a<\/p>\n<ul>\n<li>OpenCV \u2014 \u540c\u6b65\uff1a20.1 FPS<\/li>\n<li>Roboflow-\u5f02\u6b65\uff1a22.3 FPS<\/li>\n<\/ul>\n<p>\u591a\u7ebf\u7a0b\u56e0\u6b64\u5bfc\u81f4\u4e86\u9002\u5ea6\u7684\u6027\u80fd\u63d0\u5347\uff0c\u8fbe\u5230\u4e8611%\uff0c\u8868\u73b0\u76f8\u5f53\u4e0d\u9519\u3002<\/p>\n<h2>3\u3001\u9009\u62e9\u5408\u9002\u7684\u63a8\u7406\u5f15\u64ce<\/h2>\n<p>\u5728\u8bb8\u591a\u60c5\u51b5\u4e0b\uff0c\u5b9e\u65f6\u89c6\u89c9\u5e94\u7528\u7684\u6027\u80fd\u74f6\u9888\u5e76\u4e0d\u662fCPU\u6267\u884c\uff1b\u76f8\u53cd\uff0cGPU\u63a8\u7406\u65f6\u95f4\u53ef\u80fd\u4f1a\u5f88\u957f\uff0c\u5c24\u5176\u662f\u5728\u4f7f\u7528\u8f83\u5927\u6a21\u578b\u65f6\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u9009\u62e9\u6700\u9002\u5408\u76ee\u6807\u8bbe\u5907\u786c\u4ef6\u7684<strong>\u63a8\u7406\u8fd0\u884c\u65f6\u5f15\u64ce<\/strong>\u975e\u5e38\u91cd\u8981\u3002<\/p>\n<p>\u673a\u5668\u5b66\u4e60\u6a21\u578b\u901a\u5e38\u5728PyTorch\u6216TensorFlow\u7b49\u6846\u67b6\u4e2d\u8bad\u7ec3\uff0c\u8fd9\u4e9b\u6846\u67b6\u65e8\u5728\u4fc3\u8fdb\u6a21\u578b\u8bad\u7ec3\u3002\u4e00\u65e6\u8bad\u7ec3\u5b8c\u6210\uff0c\u8fd9\u4e9b\u6a21\u578b\u53ef\u4ee5\u8f6c\u6362\u4e3a\u4ee5\u52a0\u901f\u540e\u7aef\u4e3a\u7279\u70b9\u7684\u90e8\u7f72\u4e13\u7528\u683c\u5f0f\uff0c\u4ee5\u4fbf\u66f4\u5feb\u5730\u8fdb\u884c\u63a8\u7406\u3002\u8fd9\u4e9b\u63a8\u7406\u5f15\u64ce\u901a\u5e38\u9488\u5bf9\u7279\u5b9a\u7684\u82af\u7247\u7ec4\u8fdb\u884c\u4e86\u4f18\u5316\u3002\u8fd9\u4e9b\u5305\u62ec\u4f46\u4e0d\u9650\u4e8e\uff1a<\/p>\n<ul>\n<li>TensorRT \u2014 Nvidia GPU<\/li>\n<li>OpenVINO \u2014 Intel x86 CPU<\/li>\n<li>Tensorflow Lite \/ Lite Micro \u2014 \u5d4c\u5165\u5f0f<\/li>\n<li>ONNX \u2014 \u591a\u5e73\u53f0<\/li>\n<\/ul>\n<p>\u8ba9\u6211\u4eec\u56de\u5230\u4e4b\u524d\u7684\u4f8b\u5b50\u3002Roboflow\u63a8\u7406\u9ed8\u8ba4\u4f7f\u7528ONNX\u8fd0\u884c\u65f6\u5f15\u64ce\u3002\u4e3a\u4e86\u5145\u5206\u5229\u7528\u6211\u4eec\u7684Jetson\u7684Nvidia GPU\uff0c\u6211\u4eec\u53ef\u4ee5\u542f\u7528TensorRT\uff0c\u65b9\u6cd5\u662f\u8bbe\u7f6e\u4ee5\u4e0b\u73af\u5883\u53d8\u91cf\uff1a<\/p>\n<pre><code>export ONNXRUNTIME_EXECUTION_PROVIDERS=\"[TensorrtExecutionProvider,CUDAExecutionProvider,OpenVINOExecutionProvider,CoreMLExecutionProvider,CPUExecutionProvider]\"\n<\/code><\/pre>\n<p>\u8fd9\u5c06\u7f16\u8bd1\u4ece\u6211\u4eec\u7684YOLOv8\u6743\u91cd\u6587\u4ef6\u751f\u6210\u7684TensorRT\u6a21\u578b\u5f15\u64ce\uff0c\u5927\u7ea6\u9700\u898115\u5206\u949f\u3002\u518d\u6b21\u8fd0\u884c\u4e24\u4e2a\u7ba1\u9053\uff0c\u6211\u4eec\u5f97\u5230\u4ee5\u4e0b\u7ed3\u679c\uff1a<\/p>\n<ul>\n<li>OpenCV \u2014 \u540c\u6b65\uff1a26.7 FPS<\/li>\n<li>Roboflow-\u5f02\u6b65\uff1a31.9 FPS<\/li>\n<\/ul>\n<p>\u4e24\u4e2a\u7ba1\u9053\u90fd\u4eceTensorRT\u5f15\u64ce\u83b7\u5f97\u4e86\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\u3002\u6709\u8da3\u7684\u662f\uff0c\u7ba1\u9053\u4e4b\u95f4\u7684\u541e\u5410\u91cf\u5dee\u8ddd\u4ece11%\u589e\u52a0\u5230\u4e8619%\uff0c\u8fd9\u8868\u660e\u5f02\u6b65\u7ba1\u9053\u4ece\u52a0\u901f\u5f15\u64ce\u4e2d\u53d7\u76ca\u66f4\u591a\u3002<\/p>\n<h2>4\u3001\u52a0\u901f\u63a8\u7406\u7ba1\u9053\u7684\u5176\u4f59\u90e8\u5206<\/h2>\n<p>\u73b0\u5728\u6211\u4eec\u5df2\u7ecf\u52a0\u5feb\u4e86\u63a8\u7406\u6b65\u9aa4\u7684\u901f\u5ea6\uff0c\u6211\u4eec\u53ef\u4ee5\u8f6c\u5411\u589e\u5f3a\u7ba1\u9053\u7684\u5176\u4f59\u90e8\u5206\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u4f7f\u7528Nvidia\u7684Deepstream SDK\uff0c\u5b83\u5c06\u63a8\u7406\u7ba1\u9053\u62c6\u5206\u4e3a\u4e00\u7ec4\u57fa\u4e8eGStreamer\u5a92\u4f53\u6846\u67b6\u7684\u786c\u4ef6\u4f18\u5316\u63d2\u4ef6\u3002<\/p>\n<p>Deepstream\u7ba1\u9053\u3002\u6765\u6e90\uff1aNVIDIA<\/p>\n<p>\u6211\u4eec\u4f7f\u7528\u4ee5\u4e0bDeepstream\u63d2\u4ef6\u6784\u5efa\u4e86\u4e00\u4e2a\u76f8\u540c\u7684\u63a8\u7406\u7ba1\u9053\uff1a<\/p>\n<ol>\n<li>UriDecodeBin \u2014 \u81ea\u52a8\u89e3\u7801\u5404\u79cd\u89c6\u9891\u6e90\uff0c\u5305\u62ec\u4fdd\u5b58\u7684\u6587\u4ef6\u3001USB\u6444\u50cf\u5934\u548cRTSP\u6d41\u3002<\/li>\n<li>Nvstreammux \u2014 \u6279\u91cf\u5e27\u5e76\u5b9e\u4f8b\u5316NvDsBatchMeta\u4ee5\u4fdd\u5b58\u63a8\u7406\u7ed3\u679c\u548c\u5e27\u5143\u6570\u636e\u3002<\/li>\n<li>Nvinfer \u2014 TensorRT\u63a8\u7406\u5f15\u64ce\u3002<\/li>\n<li>Nvvideoconvert \u2014 \u5c06\u5e27\u8f6c\u6362\u4e3a\u53ef\u7528\u4e8e\u663e\u793a\u7684\u683c\u5f0f\u3002<\/li>\n<li>Nvosd \u2014 \u5c4f\u5e55\u663e\u793a\u5bb9\u5668\u3002<\/li>\n<li>Sink \u2014 \u5728\u684c\u9762\u4e0a\u663e\u793a\u6216\u4fdd\u5b58\u5230\u6587\u4ef6\u3002<\/li>\n<li>Queue \u2014 \u5728\u7ba1\u9053\u5143\u7d20\u4e4b\u95f4\u542f\u7528\u5f02\u6b65\u6027\u3002<\/li>\n<\/ol>\n<p>\u8fd9\u76f8\u5f53\u4e8e\u4f7f\u7528Deepstream Python\u7ed1\u5b9a\u7684\u4ee5\u4e0bPython\u5e94\u7528\u7a0b\u5e8f\uff1a<\/p>\n<pre><code>import gi  \ngi.require_version('Gst', '1.0')  \nfrom gi.repository import GLib, Gst  \nimport sys  \n# \u6dfb\u52a0\u5305\u542bdeepstream-python-apps\u4e2d\u7684\u8f85\u52a9\u51fd\u6570\u7684\u76ee\u5f55  \nsys.path.append('\/opt\/nvidia\/deepstream\/deepstream-6.3\/sources\/deepstream_python_apps\/apps')  \nfrom common.bus_call import bus_call  \nimport pyds  \n  \n# \u5e94\u7528\u7a0b\u5e8f\u7684DeepStream\u914d\u7f6e  \n# \u521d\u59cb\u5316GST  \nGst.init(None)  \n  \n# \u521b\u5efa\u7ba1\u9053  \npipeline = Gst.Pipeline()  \n  \n# \u521b\u5efa\u89c6\u9891\u6e90  \nuri_name = 'file:\/\/\/path\/to\/video'  \nsource_bin=create_source_bin(0, uri_name)  \npipeline.add(source_bin)  \n  \n# Streammux  \nstreammux = Gst.ElementFactory.make('nvstreammux', 'streammux')  \npipeline.add(streammux)  \n  \n# \u8bbe\u7f6estreammux\u5c5e\u6027  \nstreammux.set_property('batch-size', 1)  \nstreammux.set_property('width', 852)  \nstreammux.set_property('height', 480)  \n  \nsinkpad= streammux.get_request_pad(\"sink_0\")  \nsrcpad=source_bin.get_static_pad(\"src\")  \nsrcpad.link(sinkpad)   \n  \n# \u521b\u5efa\u4e3b\u8981\u63a8\u7406\u5143\u7d20\uff08GIE\uff09  \nprimary_gie = Gst.ElementFactory.make('nvinfer', 'primary-gie')  \nprimary_gie.set_property('config-file-path', '\/path\/to\/yolov8\/config\/file')  \n    \n# \u521b\u5efanvvidconv\uff08\u89c6\u9891\u8f6c\u6362\u5668\uff09  \nnvvidconv = Gst.ElementFactory.make('nvvideoconvert', 'nvvidconv')  \n  \n# \u521b\u5efaOSD\uff08\u5c4f\u5e55\u663e\u793a\uff09  \nosd = Gst.ElementFactory.make('nvdsosd', 'osd')  \n  \n# \u521b\u5efasink\uff08\u663e\u793a\u8f93\u51fa\uff09  \nsink = Gst.ElementFactory.make('nv3dsink', 'sink')  \nsink.set_property('sync', 0)  \n  \n# \u5411\u7ba1\u9053\u6dfb\u52a0\u5143\u7d20  \npipeline.add(primary_gie)  \npipeline.add(nvvidconv)  \npipeline.add(osd)  \npipeline.add(sink)  \n  \n# \u521b\u5efa\u961f\u5217\u5143\u7d20\u4ee5\u542f\u7528\u5f02\u6b65\u7ba1\u9053  \nqueue1=Gst.ElementFactory.make(\"queue\",\"queue1\")  \nqueue2=Gst.ElementFactory.make(\"queue\",\"queue2\")  \nqueue3=Gst.ElementFactory.make(\"queue\",\"queue3\")  \npipeline.add(queue1)  \npipeline.add(queue2)  \npipeline.add(queue3)  \n  \n# \u8fde\u63a5\u5143\u7d20  \nstreammux.link(queue1)  \nqueue1.link(primary_gie)  \nprimary_gie.link(queue2)  \nqueue2.link(nvvidconv)  \nnvvidconv.link(queue3)  \nqueue3.link(osd)  \nosd.link(sink)  \n  \n# \u521d\u59cb\u5316\u5faa\u73af  \nloop = GLib.MainLoop()  \nbus = pipeline.get_bus()  \nbus.add_signal_watch()  \nbus.connect (\"message\", bus_call, loop)  \n  \n# \u542f\u52a8\u7ba1\u9053  \npipeline.set_state(Gst.State.PLAYING)  \n  \n# \u542f\u52a8\u4e3b\u5faa\u73af  \ntry:    \n    loop.run()  \nexcept Exception as e:  \n    print(\"Error during pipeline execution:\", e)  \n    pass  \n  \npipeline.set_state(Gst.State.NULL)\n<\/code><\/pre>\n<p>\u6211\u4eec\u4f7f\u7528\u4ed3\u5e93\u5c06\u6211\u4eec\u7684PyTorch\u8f6c\u6362\u4e3aTensorRT\u6a21\u578b\u5f15\u64ce\u6587\u4ef6\u3002\u8fd0\u884c\u8be5\u5e94\u7528\u7a0b\u5e8f\uff0c\u6211\u4eec\u5f97\u5230\u4ee5\u4e0b\u541e\u5410\u91cf\u6d4b\u91cf\u7ed3\u679c\uff1a<\/p>\n<ul>\n<li>Deepstream-\u5f02\u6b65\uff1a48.2 FPS<\/li>\n<\/ul>\n<p>\u8fd9\u8868\u660e\u76f8\u5bf9\u4e8e\u524d\u9762\u4e24\u79cd\u5b9e\u73b0\uff0c\u6027\u80fd\u53c8\u6709\u4e86\u663e\u8457\u63d0\u5347\u3002<\/p>\n<h2><strong>5\u3001\u7ed3\u675f\u8bed<\/strong><\/h2>\n<p>\u603b\u4e4b\uff0c\u672c\u6559\u7a0b\u6f14\u793a\u4e86\u5982\u4f55\u4f7f\u7528Roboflow\u63a8\u7406\u7ba1\u9053\u548cDeepstream SDK\u5b9e\u73b0\u5f02\u6b65\u548c\u52a0\u901f\u63a8\u7406\u7ba1\u9053\uff0c\u4f7f\u541e\u5410\u91cf\u63d0\u9ad8\u4e86\u4e24\u500d\u4ee5\u4e0a\u3002<\/p>\n<p>\u5bf9\u4e8e\u4e25\u91cd\u53d7\u9650\u7684\u8bbe\u5907\uff0c\u8fd8\u53ef\u4ee5\u901a\u8fc7\u5176\u4ed6\u6280\u672f\u8fdb\u4e00\u6b65\u51cf\u5c11\u63a8\u7406\u5ef6\u8fdf\uff0c\u5305\u62ec\u5e27\u6279\u5904\u7406\u3001\u6a21\u578b\u91cf\u5316\u548c\u526a\u679d\u4ee5\u53ca\u4f7f\u7528\u66f4\u7b80\u5355\u7684\u6a21\u578b\u3002\u5e0c\u671b\u8fd9\u7bc7\u6587\u7ae0\u80fd\u5e2e\u52a9\u4f60\u5f00\u53d1\u5b9e\u65f6\u89c6\u89c9\u5e94\u7528\u3002<\/p>\n<p><em>\u8be5\u9879\u76ee\u7684\u5b8c\u6574\u4ee3\u7801\u53ef\u4ee5\u5728\u8fd9\u91cc\u627e\u5230<\/em><em>\u3002<\/em><\/p>\n<hr>\n","protected":false},"excerpt":{"rendered":"<p>\u5728\u8fb9\u7f18\u90e8\u7f72\u673a\u5668\u5b66\u4e60\u901a\u5e38\u4f1a\u5728\u4f20\u611f\u548c\u76d1\u63a7\u5e94\u7528\u4e2d\u63d0\u4f9b\u663e\u8457\u4f18\u52bf\u3002\u901a\u8fc7\u5c06\u5904\u7406\u4fdd\u6301\u5728\u6570\u636e\u6e90\u9644\u8fd1\uff0c\u53ef\u4ee5\u8282\u7701\u5927\u91cf\u7684\u7f51\u7edc\u3001\u5b58\u50a8\u548c\u4e91\u8ba1\u7b97\u6210\u672c\uff0c\u5e76\u4e14\u6574\u4f53\u5ef6\u8fdf\u4e5f\u4f1a\u964d\u4f4e\u3002\u7136\u800c\uff0c\u8fb9\u7f18\u8bbe\u5907\u53d7\u9650\u4e8e\u5176\u6709\u9650\u7684\u8ba1\u7b97\u8d44\u6e90\uff0c\u8fd9\u4e9b\u8d44\u6e90\u901a\u5e38\u6bd4\u670d\u52a1\u5668\u7684\u8981\u5f31\u3002\u56e0\u6b64\uff0c\u5f53\u5728\u8fb9\u7f18\u90e8\u7f72\u673a\u5668\u5b66\u4e60\u5e94\u7528\u7a0b\u5e8f\u65f6\uff0c\u5fc5\u987b\u5bf9\u5176\u8fdb\u884c\u4e13\u95e8\u4f18\u5316\u4ee5\u6700\u5927\u5316\u6027\u80fd\u3002 \u5728\u8fd9\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u5c06\u8ba8\u8bba\u52a0\u901f\u5b9e\u65f6\u8ba1\u7b97\u673a\u89c6\u89c9\u5e94\u7528\u7684\u6280\u672f\uff0c\u8fd9\u4e9b\u6280\u672f\u53ef\u4ee5\u5728\u667a\u80fd\u4f20\u611f\u3001\u76d1\u63a7\u6216\u673a\u5668\u4eba\u7b49\u573a\u666f\u4e2d\u627e\u5230\u3002\u6211\u4eec\u5c06\u901a\u8fc7\u4f7f\u7528\u5f00\u6e90\u5e93\uff08\u5982OpenCV\u3001Roboflow\u63a8\u7406\u548cNvidia Deepstream\uff09\u6765\u5b9e\u73b0\u63a8\u7406\u7ba1\u9053\u3002 1\u3001\u63a8\u7406\u7ba1\u9053 \u63a8\u7406\u7ba1\u9053\u5c06\u8fd0\u884c\u89c6\u9891\u6d41\u4e0a\u7684\u673a\u5668\u5b66\u4e60\u63a8\u7406\u7684\u8fc7\u7a0b\u5206\u4e3a\u4e00\u7cfb\u5217\u79bb\u6563\u6b65\u9aa4\u3002\u4e00\u4e2a\u57fa\u672c\u7684\u7ba1\u9053\u901a\u5e38\u5305\u62ec\u4ee5\u4e0b\u9636\u6bb5\uff1a \u89c6\u9891\u6d41\u6570\u636e\u88ab\u63d0\u53d6\u5e76\u89e3\u7801\u5f62\u6210\u56fe\u50cf\u5e27\u3002 \u5e27\u5728\u63a8\u7406\u524d\u8fdb\u884c\u9884\u5904\u7406\uff08\u8c03\u6574\u5927\u5c0f\u3001\u5f52\u4e00\u5316\u7b49\uff09\u3002 \u6279\u91cf\u5e27\u88ab\u8f6c\u6362\u4e3a\u5f20\u91cf\u5e76\u53d1\u9001\u5230\u8bbe\u5907GPU\uff08\u5982\u679c\u6ca1\u6709\u53ef\u7528GPU\uff0c\u5219\u53d1\u9001\u5230CPU\uff09\u8fdb\u884c\u63a8\u7406\u3002 \u63a8\u7406\u7ed3\u679c\u88ab\u53e0\u52a0\u5230\u539f\u59cb\u5e27\u4e0a\uff0c\u5e76\u663e\u793a\u5728\u5c4f\u5e55\u4e0a\u3001\u672c\u5730\u4fdd\u5b58\u6216\u5411\u4e0b\u4f20\u8f93\u3002 \u63a8\u7406\u7ed3\u679c\u4f5c\u4e3a\u5143\u6570\u636e\u53d1\u9001\u4ee5\u89e6\u53d1\u8fdb\u4e00\u6b65\u64cd\u4f5c\u3002 \u66f4\u590d\u6742\u7684\u7ba1\u9053\u53ef\u80fd\u5305\u62ec\u5bf9\u8c61\u8ddf\u8e2a\u3001\u611f\u5174\u8da3\u533a\u57df\u8fc7\u6ee4\u548c\u5176\u4ed6\u4e1a\u52a1\u903b\u8f91\uff0c\u5177\u4f53\u53d6\u51b3\u4e8e\u5e94\u7528\u573a\u666f\u3002 2\u3001\u63a8\u7406\u7ba1\u9053\u7684\u5f02\u6b65\u5316 \u867d\u7136\u63a8\u7406\u7ba1\u9053\u4e2d\u7684\u6b65\u9aa4\u662f\u4f9d\u6b21\u6267\u884c\u7684\uff0c\u4f46\u5b83\u4eec\u5b9e\u9645\u4e0a\u53ef\u4ee5\u901a\u8fc7\u4f7f\u7528\u591a\u7ebf\u7a0b\u5728\u8fd0\u884c\u65f6\u5e76\u53d1\u6267\u884c\u3002 \u591a\u7ebf\u7a0b\u662f\u6307\u5728\u540c\u4e00\u65f6\u95f4\u8fd0\u884c\u591a\u4e2a\u7ebf\u7a0b\u6216\u53ef\u8fd0\u884c\u4ee3\u7801\u6bb5\u7684\u8fc7\u7a0b\u3002\u5728Python\u4e2d\uff0c\u8fd9\u662f\u901a\u8fc7\u4e0a\u4e0b\u6587\u5207\u6362\u5b9e\u73b0\u7684\uff0c\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u5904\u7406\u5668\u5728\u5f53\u524d\u7ebf\u7a0b\u7a7a\u95f2\u65f6\u5207\u6362\u5230\u5176\u4ed6\u7ebf\u7a0b\uff0c\u4ece\u800c\u7ed9\u4eba\u4e00\u79cd\u4e24\u8005\u540c\u65f6\u6267\u884c\u7684\u9519\u89c9\u3002 \u7531\u4e8e\u63a8\u7406\u7ba1\u9053\u5305\u62ecI\/O\u64cd\u4f5c\uff08\u8bfb\u53d6\u5e27\u3001\u5c06\u5f20\u91cf\u590d\u5236\u5230GPU\u3001\u4f20\u8f93\u7ed3\u679c\uff09\uff0c\u5728\u6b64\u671f\u95f4CPU\u5904\u4e8e\u7b49\u5f85\u72b6\u6001\uff0c\u56e0\u6b64\u591a\u7ebf\u7a0b\u63d0\u4f9b\u4e86\u52a0\u901f\u7684\u6f5c\u529b\u3002\u8fd9\u5728\u5e27\u4ee5\u6279\u6b21\u65b9\u5f0f\u5904\u7406\u65f6\u7279\u522b\u6709\u7528\uff0c\u56e0\u4e3a\u8fd9\u53ef\u4ee5\u8ba9CPU\u6709\u65f6\u95f4\u7ee7\u7eed\u5904\u7406\u5176\u4ed6\u4efb\u52a1\uff0c\u800c\u5f53\u6279\u6b21\u6b63\u5728\u6536\u96c6\u548c\u53d1\u9001\u8fdb\u884c\u63a8\u7406\u65f6\u3002 \u8fd9\u5c31\u662fRoboflow\u7684\u63a8\u7406\u7ba1\u9053\u5b9e\u73b0\u7684\u57fa\u7840\uff0c\u5b83\u5c06\u89c6\u9891\u89e3\u7801\u3001\u63a8\u7406\u548c\u540e\u5904\u7406\u5206\u79bb\u5230\u4e13\u7528\u7ebf\u7a0b\u4e2d\u5e76\u53d1\u8fd0\u884c\u3002 \u591a\u7ebf\u7a0b\u7ba1\u9053\u3002\u6765\u6e90\uff1a \u8003\u8651\u4ee5\u4e0b\u4f7f\u7528OpenCV\u4e2d\u7684YOLOv8-nano\u76ee\u6807\u68c0\u6d4b\u6a21\u578b\u5b9e\u73b0\u7684\u540c\u6b65\u63a8\u7406\u7ba1\u9053\uff1a import cv2 from inference import get_model import supervision as sv model = get_model(&#8220;yolov8n-640&#8221;) cap = cv2.VideoCapture(&#8216;\/path\/to\/video&#8217;) while cap.isOpened(): ret, frame = cap.read() if not ret: print(&#8220;Can&#8217;t receive frame (stream end?). Exiting &#8230;&#8221;) break predictions = model.infer(frame) detections [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[13],"tags":[],"class_list":["post-53724","post","type-post","status-publish","format-standard","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53724","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/comments?post=53724"}],"version-history":[{"count":0,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53724\/revisions"}],"wp:attachment":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/media?parent=53724"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/categories?post=53724"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/tags?post=53724"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}