RF-DETR SORT 多目标跟踪实战轻量级跟踪器快速上手这篇教程根据我复现 SORT 多目标跟踪流程时整理重点演示命令行跟踪、Python 代码跟踪和运动补偿扩展。本文整理自我的学习和项目复现过程尽量按实操顺序保留 notebook 的关键步骤同时把数据集获取方式调整为适合中文教程发布的写法。本文会重点跑通以下流程安装跟踪依赖准备示例视频或自己的视频使用 CLI 运行 SORT 跟踪用 Python 组合检测器和 SORTTracker在滑雪等镜头运动场景加入运动补偿如果你正在系统学习目标检测、实例分割、OCR、多目标跟踪或视觉大模型建议收藏本文配套 notebook、示例图片和运行环境说明后续会继续整理。如果环境配置卡住可以在评论区说明具体报错。 文章目录RF-DETR SORT 多目标跟踪实战轻量级跟踪器快速上手⚙️ 环境准备 准备视频 命令行跟踪 Python 跟踪流程 运动补偿跟踪 小结 同系列教程汇总⚙️ 环境准备先检查 GPU 并安装跟踪相关依赖。!nvidia-smi!pip install-q inference-gpu trackers2.3.0 准备视频下载示例视频也可以替换为自己的本地视频。# 准备示例视频。可以把自己的视频上传到 /content 后按下面文件名命名或同步修改后续路径。SOURCE_VIDEO_1/content/bikes-1280x720-1.mp4SOURCE_VIDEO_2/content/bikes-1280x720-2.mp4SOURCE_VIDEO_3/content/skiers-1280x720-5.mp4print(video placeholders ready) 命令行跟踪使用 trackers CLI 快速跑通 SORT 跟踪流程。SOURCE_VIDEO_PATH/content/bikes-1280x720-1.mp4TARGET_VIDEO_PATH/content/bikes-1280x720-1-result.mp4!trackers track \--source{SOURCE_VIDEO_PATH}\--output{TARGET_VIDEO_PATH}\--model rfdetr-medium \--tracker sort \--show-trajectoriesTARGET_VIDEO_COMPRESSED_PATH/content/bikes-1280x720-1-result-compressed.mp4!ffmpeg-y-loglevel error-i{TARGET_VIDEO_PATH}-vcodec libx264-crf28{TARGET_VIDEO_COMPRESSED_PATH}fromIPython.displayimportVideo Video(TARGET_VIDEO_COMPRESSED_PATH,embedTrue,width1080) Python 跟踪流程用 Python 代码组合检测模型和 SORTTracker。frominferenceimportget_modelfromtrackersimportSORTTracker modelget_model(rfdetr-medium)trackerSORTTracker()importsupervisionassv colorsv.ColorPalette.from_hex([#ffff00,#ff9b00,#ff8080,#ff66b2,#ff66ff,#b266ff,#9999ff,#3399ff,#66ffff,#33ff99,#66ff66,#99ff00])box_annotatorsv.BoxAnnotator(colorcolor,color_lookupsv.ColorLookup.TRACK)label_annotatorsv.LabelAnnotator(colorcolor,color_lookupsv.ColorLookup.TRACK,text_colorsv.Color.BLACK,text_scale0.8)trace_annotatorsv.TraceAnnotator(colorcolor,color_lookupsv.ColorLookup.TRACK,thickness2,trace_length100)CONFIDENCE_THRESHOLD0.2NMS_THRESHOLD0.3SOURCE_VIDEO_PATH/content/bikes-1280x720-2.mp4TARGET_VIDEO_PATH/content/bikes-1280x720-2-result.mp4defcallback(frame,i):resultmodel.infer(frame,confidenceCONFIDENCE_THRESHOLD)[0]detectionssv.Detections.from_inference(result).with_nms(thresholdNMS_THRESHOLD)detectionstracker.update(detections)annotated_imageframe.copy()annotated_imagebox_annotator.annotate(annotated_image,detections)annotated_imagetrace_annotator.annotate(annotated_image,detections)annotated_imagelabel_annotator.annotate(annotated_image,detections,detections.tracker_id)returnannotated_image tracker.reset()sv.process_video(source_pathSOURCE_VIDEO_PATH,target_pathTARGET_VIDEO_PATH,callbackcallback,show_progressTrue,)TARGET_VIDEO_COMPRESSED_PATH/content/bikes-1280x720-2-result-compressed.mp4!ffmpeg-y-loglevel error-i{TARGET_VIDEO_PATH}-vcodec libx264-crf28{TARGET_VIDEO_COMPRESSED_PATH}fromIPython.displayimportVideo Video(TARGET_VIDEO_COMPRESSED_PATH,embedTrue,width1080) 运动补偿跟踪在镜头运动明显的场景中加入运动估计和轨迹补偿。frominferenceimportget_modelfromtrackersimportSORTTracker,MotionEstimator,MotionAwareTraceAnnotator PERSON_CLASS_ID0modelget_model(rfdetr-large)trackerSORTTracker(minimum_consecutive_frames3)motion_estimatorMotionEstimator(max_points500,min_distance10,quality_level0.001,ransac_reproj_threshold1.0,)colorsv.ColorPalette.from_hex([#ffff00,#ff9b00,#ff8080,#ff66b2,#ff66ff,#b266ff,#9999ff,#3399ff,#66ffff,#33ff99,#66ff66,#99ff00])box_annotatorsv.BoxAnnotator(colorcolor,color_lookupsv.ColorLookup.TRACK)label_annotatorsv.LabelAnnotator(colorcolor,color_lookupsv.ColorLookup.TRACK,text_colorsv.Color.BLACK,text_scale0.8)motion_aware_trace_annotatorMotionAwareTraceAnnotator(colorcolor,color_lookupsv.ColorLookup.TRACK,thickness2,trace_length100)CONFIDENCE_THRESHOLD0.2NMS_THRESHOLD0.3SOURCE_VIDEO_PATH/content/skiers-1280x720-5.mp4TARGET_VIDEO_PATH/content/skiers-1280x720-5-result.mp4defcallback(frame,i):coord_transformmotion_estimator.update(frame)resultmodel.infer(frame,confidenceCONFIDENCE_THRESHOLD)[0]detectionssv.Detections.from_inference(result).with_nms(thresholdNMS_THRESHOLD)detectionsdetections[detections.class_idPERSON_CLASS_ID]detectionstracker.update(detections)annotated_imageframe.copy()annotated_imagebox_annotator.annotate(annotated_image,detections)annotated_imagemotion_aware_trace_annotator.annotate(annotated_image,detections,coord_transformcoord_transform)annotated_imagelabel_annotator.annotate(annotated_image,detections,detections.tracker_id)returnannotated_image tracker.reset()motion_estimator.reset()motion_aware_trace_annotator.reset()sv.process_video(source_pathSOURCE_VIDEO_PATH,target_pathTARGET_VIDEO_PATH,callbackcallback,show_progressTrue,)TARGET_VIDEO_COMPRESSED_PATH/content/skiers-1280x720-5-result-compressed.mp4!ffmpeg-y-loglevel error-i{TARGET_VIDEO_PATH}-vcodec libx264-crf28{TARGET_VIDEO_COMPRESSED_PATH}fromIPython.displayimportVideo Video(TARGET_VIDEO_COMPRESSED_PATH,embedTrue,width1080) 小结这篇教程完整整理了RF-DETR 与 SORT 多目标跟踪的核心复现流程。实际操作时建议先确认 GPU、依赖版本、数据集路径和模型权重路径再逐段运行 notebook。后续我会继续按源项目顺序整理同系列中的目标检测、实例分割、OCR、多目标跟踪和视觉大模型教程。 同系列教程汇总Google Gemini 3.5 Flash 零样本目标检测教程从提示词到可视化结果GLM-OCR 文档识别实战教程从验证码、公式到车牌 OCRRF-DETR ByteTrack 多目标跟踪实战教程从命令行到 Python 视频轨迹可视化SAM 3 图像分割实战教程文本、框和点提示的多种分割方式-RF-DETR SORT 多目标跟踪实战轻量级跟踪器快速上手