RF-DETR + OC-SORT 多目标跟踪实战：遮挡场景下的轨迹稳定性优化
RF-DETR + OC-SORT 多目标跟踪实战：遮挡场景下的轨迹稳定性优化
这篇教程根据我复现 OC-SORT 多目标跟踪流程时整理，重点演示快速跟踪、Python 回调处理和复杂运动场景下的轨迹增强。
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 OC-SORT 跟踪
- 用 Python 组合检测器和 OCSORTTracker
- 结合运动补偿处理镜头运动场景
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 + OC-SORT 多目标跟踪实战：遮挡场景下的轨迹稳定性优化
- ⚙️ 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 trackers==2.3.0🎬 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.mp4"SOURCE_VIDEO_2="/content/bikes-1280x720-2.mp4"SOURCE_VIDEO_3="/content/skiers-1280x720-5.mp4"print("video placeholders ready")🚀 命令行跟踪
使用 trackers CLI 快速跑通 OC-SORT 跟踪流程。
SOURCE_VIDEO_PATH="/content/bikes-1280x720-1.mp4"TARGET_VIDEO_PATH="/content/bikes-1280x720-1-result.mp4"!trackers track \--source{SOURCE_VIDEO_PATH}\--output{TARGET_VIDEO_PATH}\--model rfdetr-medium \--tracker ocsort \--show_trajectories trueTARGET_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,embed=True,width=1080)🐍 Python 跟踪流程
用 Python 代码组合检测模型和 OCSORTTracker。
frominferenceimportget_modelfromtrackersimportOCSORTTracker model=get_model("rfdetr-medium")tracker=OCSORTTracker()importsupervisionassv color=sv.ColorPalette.from_hex(["#ffff00","#ff9b00","#ff8080","#ff66b2","#ff66ff","#b266ff","#9999ff","#3399ff","#66ffff","#33ff99","#66ff66","#99ff00"])box_annotator=sv.BoxAnnotator(color=color,color_lookup=sv.ColorLookup.TRACK)label_annotator=sv.LabelAnnotator(color=color,color_lookup=sv.ColorLookup.TRACK,text_color=sv.Color.BLACK,text_scale=0.8)trace_annotator=sv.TraceAnnotator(color=color,color_lookup=sv.ColorLookup.TRACK,thickness=2,trace_length=100)CONFIDENCE_THRESHOLD=0.2NMS_THRESHOLD=0.3SOURCE_VIDEO_PATH="/content/bikes-1280x720-2.mp4"TARGET_VIDEO_PATH="/content/bikes-1280x720-2-result.mp4"defcallback(frame,i):result=model.infer(frame,confidence=CONFIDENCE_THRESHOLD)[0]detections=sv.Detections.from_inference(result).with_nms(threshold=NMS_THRESHOLD)detections=tracker.update(detections)annotated_image=frame.copy()annotated_image=box_annotator.annotate(annotated_image,detections)annotated_image=trace_annotator.annotate(annotated_image,detections)annotated_image=label_annotator.annotate(annotated_image,detections,detections.tracker_id)returnannotated_image tracker.reset()sv.process_video(source_path=SOURCE_VIDEO_PATH,target_path=TARGET_VIDEO_PATH,callback=callback,show_progress=True,)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,embed=True,width=1080)🧭 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_modelfromtrackersimportOCSORTTracker,MotionEstimator,MotionAwareTraceAnnotator PERSON_CLASS_ID=0model=get_model("rfdetr-large")tracker=OCSORTTracker(minimum_consecutive_frames=3)motion_estimator=MotionEstimator(max_points=500,min_distance=10,quality_level=0.001,ransac_reproj_threshold=1.0,)color=sv.ColorPalette.from_hex(["#ffff00","#ff9b00","#ff8080","#ff66b2","#ff66ff","#b266ff","#9999ff","#3399ff","#66ffff","#33ff99","#66ff66","#99ff00"])box_annotator=sv.BoxAnnotator(color=color,color_lookup=sv.ColorLookup.TRACK)label_annotator=sv.LabelAnnotator(color=color,color_lookup=sv.ColorLookup.TRACK,text_color=sv.Color.BLACK,text_scale=0.8)motion_aware_trace_annotator=MotionAwareTraceAnnotator(color=color,color_lookup=sv.ColorLookup.TRACK,thickness=2,trace_length=100)CONFIDENCE_THRESHOLD=0.2NMS_THRESHOLD=0.3SOURCE_VIDEO_PATH="/content/skiers-1280x720-5.mp4"TARGET_VIDEO_PATH="/content/skiers-1280x720-5-result.mp4"defcallback(frame,i):coord_transform=motion_estimator.update(frame)result=model.infer(frame,confidence=CONFIDENCE_THRESHOLD)[0]detections=sv.Detections.from_inference(result).with_nms(threshold=NMS_THRESHOLD)detections=detections[detections.class_id==PERSON_CLASS_ID]detections=tracker.update(detections)annotated_image=frame.copy()annotated_image=box_annotator.annotate(annotated_image,detections)annotated_image=motion_aware_trace_annotator.annotate(annotated_image,detections,coord_transform=coord_transform)annotated_image=label_annotator.annotate(annotated_image,detections,detections.tracker_id)returnannotated_image tracker.reset()motion_estimator.reset()motion_aware_trace_annotator.reset()sv.process_video(source_path=SOURCE_VIDEO_PATH,target_path=TARGET_VIDEO_PATH,callback=callback,show_progress=True,)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,embed=True,width=1080)📌 小结
这篇教程完整整理了RF-DETR 与 OC-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 OCR
RF-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 + OC-SORT 多目标跟踪实战：遮挡场景下的轨迹稳定性优化