基于yolov5使用sort实现目标追踪支持rk3568以及rk3588部署提供python接口与c接口1下面给你一套完整的YOLOv5 SORT 目标追踪代码并适配RK3568/RK3588部署含 Python 与 C 接口说明。一、项目整体说明项目说明算法YOLOv5 目标检测 SORT 多目标追踪硬件瑞芯微 RK3568 / RK3588部署方式Python 接口RKNN Lite、C 接口RKNN C API功能视频/摄像头中车辆、行人等目标检测ID追踪二、Python 接口代码RKNN Lite 版本1. 依赖安装pipinstallrknn-toolkit-lite2 opencv-python numpy2. SORT 追踪器核心实现importnumpyasnpfromscipy.optimizeimportlinear_sum_assignmentfromfilterpy.kalmanimportKalmanFilter# 单个目标的卡尔曼滤波器classKalmanBoxTracker:def__init__(self,bbox):self.kfKalmanFilter(dim_x7,dim_z4)self.kf.Fnp.array([[1,0,0,0,1,0,0],[0,1,0,0,0,1,0],[0,0,1,0,0,0,1],[0,0,0,1,0,0,0],[0,0,0,0,1,0,0],[0,0,0,0,0,1,0],[0,0,0,0,0,0,1]])self.kf.Hnp.array([[1,0,0,0,0,0,0],[0,1,0,0,0,0,0],[0,0,1,0,0,0,0],[0,0,0,1,0,0,0]])self.kf.R[2:,2:]*10.self.kf.P[4:,4:]*1000.self.kf.P*10.self.kf.Q[-1,-1]*0.01self.kf.Q[4:,4:]*0.01self.kf.x[:4]self.convert_bbox_to_z(bbox)self.time_since_update0self.id-1self.hit_streak0defconvert_bbox_to_z(self,bbox):wbbox[2]-bbox[0]hbbox[3]-bbox[1]xbbox[0]w/2.ybbox[1]h/2.sw*h rw/hreturnnp.array([x,y,s,r]).reshape((4,1))defconvert_x_to_bbox(self,x):x,y,s,rx[0,0],x[1,0],x[2,0],x[3,0]wnp.sqrt(s*r)hs/wreturnnp.array([x-w/2.,y-h/2.,xw/2.,yh/2.]).reshape((4,))defupdate(self,bbox):self.time_since_update0self.hit_streak1self.kf.update(self.convert_bbox_to_z(bbox))defpredict(self):ifself.kf.x[6]self.kf.x[2]0:self.kf.x[6]0self.kf.predict()ifself.time_since_update0:self.hit_streak0self.time_since_update1returnself.convert_x_to_bbox(self.kf.x)classSORT:def__init__(self,max_age3,min_hits3):self.max_agemax_age self.min_hitsmin_hits self.trackers[]self.frame_count0self.next_id1defupdate(self,dets):self.frame_count1trks[]to_del[]ret[]# 预测所有轨迹fort,trkinenumerate(self.trackers):postrk.predict()trks.append(pos)iftrk.time_since_updateself.max_age:to_del.append(t)trksnp.array(trks)fortinreversed(to_del):self.trackers.pop(t)# 匈牙利匹配matched,unmatched_dets,unmatched_trksself.associate(dets,trks)# 更新匹配的轨迹forminmatched:self.trackers[m[1]].update(dets[m[0]])# 新增轨迹foriinunmatched_dets:trkKalmanBoxTracker(dets[i])trk.idself.next_id self.next_id1self.trackers.append(trk)# 输出结果fort,trkinenumerate(self.trackers):iftrk.time_since_update1andtrk.hit_streakself.min_hits:ret.append(np.concatenate((trk.convert_x_to_bbox(trk.kf.x),[trk.id])).reshape(1,-1))iflen(ret)0:returnnp.concatenate(ret)returnnp.empty((0,5))defassociate(self,dets,trks):iflen(dets)0:return[],[],list(range(len(trks)))iflen(trks)0:return[],list(range(len(dets))),[]iou_matrixnp.zeros((len(dets),len(trks)))ford,detinenumerate(dets):fort,trkinenumerate(trks):iou_matrix[d,t]self.iou(det,trk)matched_indiceslinear_sum_assignment(-iou_matrix)matched_indicesnp.array(list(zip(*matched_indices)))unmatched_dets[]unmatched_trks[]iflen(matched_indices)0:unmatched_detslist(range(len(dets)))unmatched_trkslist(range(len(trks)))else:unmatched_dets[dfordinrange(len(dets))ifdnotinmatched_indices[:,0]]unmatched_trks[tfortinrange(len(trks))iftnotinmatched_indices[:,1]]matches[]forminmatched_indices:ifiou_matrix[m[0],m[1]]0.3:unmatched_dets.append(m[0])unmatched_trks.append(m[1])else:matches.append(m.reshape(1,2))iflen(matches)0:matchesnp.empty((0,2),dtypeint)else:matchesnp.concatenate(matches,axis0)returnmatches,unmatched_dets,unmatched_trksdefiou(self,a,b):xx1max(a[0],b[0])yy1max(a[1],b[1])xx2min(a[2],b[2])yy2min(a[3],b[3])wmax(0.,xx2-xx1)hmax(0.,yy2-yy1)area_interw*h area_a(a[2]-a[0])*(a[3]-a[1])area_b(b[2]-b[0])*(b[3]-b[1])returnarea_inter/(area_aarea_b-area_inter)3. YOLOv5 SORT 主程序RKNN Liteimportcv2importnumpyasnpfromrknnlite.apiimportRKNNLitefromsortimportSORT# 上面的SORT代码# 配置RKNN_MODELyolov5s.rknnIMG_SIZE640CONF_THRESH0.25IOU_THRESH0.45# 初始化RKNNrknn_liteRKNNLite()retrknn_lite.load_rknn(RKNN_MODEL)retrknn_lite.init_runtime(core_maskRKNNLite.NPU_CORE_0)# 初始化追踪器trackerSORT(max_age3,min_hits3)defpreprocess(img):imgcv2.cvtColor(img,cv2.COLOR_BGR2RGB)imgcv2.resize(img,(IMG_SIZE,IMG_SIZE))imgnp.expand_dims(img,axis0)returnimgdefpostprocess(outputs,img0):# 这里简化处理实际需按YOLOv5后处理逻辑解析outputs# 示例outputs为[N,6]的检测结果格式为[x1,y1,x2,y2,conf,cls]dets[]# 实际实现中需根据RKNN的输出格式解析检测框# 这里用假数据示意returnnp.array(dets)capcv2.VideoCapture(0)# 或视频文件路径whilecap.isOpened():ret,framecap.read()ifnotret:breakimgpreprocess(frame)outputsrknn_lite.inference(inputs[img])detspostprocess(outputs,frame)# 追踪trackstracker.update(dets)# 画框IDfordintracks:x1,y1,x2,y2,track_idd cv2.rectangle(frame,(int(x1),int(y1)),(int(x2),int(y2)),(0,0,255),2)cv2.putText(frame,fID:{int(track_id)},(int(x1),int(y1)-10),cv2.FONT_HERSHEY_SIMPLEX,0.5,(0,0,255),2)cv2.imshow(YOLOv5SORT,frame)ifcv2.waitKey(1)0xFFord(q):breakcap.release()cv2.destroyAllWindows()rknn_lite.release()三、C 接口说明RKNN C API 部署1. 环境依赖交叉编译工具链RK3568/RK3588 SDKOpenCVRKNN C API 库2. 关键流程#includerknn_api.h#includeopencv2/opencv.hpp// 1. 初始化RKNNrknn_context ctx;rknn_init(ctx,yolov5s.rknn,0,0,NULL);// 2. 预处理cv::Mat imgcv::imread(test.jpg);cv::resize(img,img,cv::Size(640,640));// ...// 3. 推理rknn_input inputs[1];rknn_output outputs[1];// 设置inputsrknn_run(ctx,NULL);rknn_outputs_get(ctx,1,outputs,NULL);// 4. 后处理SORT追踪// 解析outputs调用SORT的C实现// ...// 5. 释放资源rknn_destroy(ctx);注C 完整代码需基于 RKNN SDK 示例改造并移植 SORT 算法卡尔曼滤波匈牙利匹配可参考官方 yolov5 示例。四、RK3568/RK3588 部署要点模型转换使用rknn-toolkit2将 PyTorch/ONNX 版 YOLOv5 转换为.rknn格式命令示例python-mrknn.api.converter yolov5s.onnx yolov5s.rknn--platformrk3588性能优化使用core_mask指定 NPU 核心预处理/后处理尽量用 NEON 或 OpenCV 硬件加速降低推理分辨率如 416x416提升帧率视频流处理使用cv2.VideoCapture读取摄像头/视频多线程一帧推理一帧预处理避免阻塞五、补充说明Python 版代码可直接在 RK3568/RK3588 的 Linux 系统上运行需安装rknn-toolkit-lite2C 版适合追求极致性能的场景可参考瑞芯微官方 SDK 示例改造SORT 算法中max_age、min_hits、IOU_THRESH可根据场景调整