online_learning 流程

[Durant35]

• online_learning 流程

  • Region-based分割得到备选的human cloud和对应的pose
  • 简单的：A human-like volumetric model is used to filter out over- and under-segmented clusters，即作为Detector结果输出给people tracker
    • 未通过human-like volumetric model的Segments直接作为负样本，不作为Detector检测结果
  • 在需要训练的阶段，train_round_ < max_trains_，Detector的结果pose才会输出给people tracker，后者维护human轨迹，当满足以下条件时publish回Detector
    • Generate trajectories over a given time interval K
    • Generate trajectories that have idled
    • 产生的轨迹会经过PN_experts，以判定轨迹是"human_trajectory"还是“static_trajectory”
      • the target moves a minimum distance within a given time interval
      • the target’s velocity is non-zero but also not faster than a person’s preferred walking speed of 1.4 m/s
      • filters out objects (true negatives) that are associated to human-like trajectories but are too “uncertain” because moving in an unexpected way or affected by the proximity of other clusters
    • 每次publish回Detector的都是一条轨迹，一条一条publish
  • 也只有在需要训练的阶段，这些轨迹才会被Detector用于辅助产生P/N samples
    • 轨迹源自Detector的多次观测，这些观测一一对应Segment：learnable_clusters_[i]，对于Classified as "human" by Human Classifier，利用轨迹来产生P/N samples，纠正SVM Human Classifier
    • True Classified --> Positive Sample
    • False Classified --> Negative Sample
    • 当P/N samples满足一定数目时max_positives_/max_negatives_，利用这个训练周期产生的样本重训练SVM Human Classifier
  • 在不需要训练的阶段，Detector的结果pose并不会输出给people tracker，后者相当于停止工作了；Detector也不会再收集P/N samples和重训练SVM Human Classifier；Detector的工作只剩下Region-based分割，通过human-like volumetric model筛选，然后通过SVM Human Classifier检测出Human

[Durant35]

object3d_detector 非 online_learning版本

• 使用训练好的 SVM 分类器

  <!-- FLOBOT 3D Object Detector -->
  <node pkg="object3d_detector" type="object3d_detector" name="object3d_detector" output="screen" respawn="false">
  <param name="model_file_name" type="string" value="$(find object3d_detector)/libsvm/pedestrian.model"/>
  <param name="range_file_name" type="string" value="$(find object3d_detector)/libsvm/pedestrian.range"/>
  </node>

• 效果似乎没有 online_learning 版本效果好
• 训练好的 Human Classifier，Human Detection 的流程变成先分割，提取Segment特征，SVM 分类器分类出 Human Segment；Detector 和 Tracker 没有feedback

[Durant35]

object3d_detector_ol online_learning版本

• 最初以为一开始需要加载线下训练好的模型，实际上并没有；系统从一开始就收集正/负样本，之后进行预先设定好的几轮训练
• 每轮训练需要准备好足够的positives/negatives样本，训练是叠加之前的所有样本进行重新训练

• P/N samples collection in Object3dDetector::trajectoryCallback() 结合Tracking轨迹

  //TODO -- True Classified --> Positive Sample
  if (trajectory->header.frame_id == "human_trajectory" && positive_ < max_positives_) {
  svm_problem_.y[svm_problem_.l++] = 1; // 1, the positive label
  ++positive_ >= max_positives_ ? stop = true : stop = false;
  //learned_cloud += *learnable_clusters_[j];
  }
  //TODO -- False Classified --> Negative Sample
  //TODO Why need "train_round_ > 0"
  if (train_round_ > 0 && trajectory->header.frame_id == "static_trajectory" &&
  negative_ < max_negatives_) {
  svm_problem_.y[svm_problem_.l++] = -1; // -1, the negative label
  ++negative_ >= max_negatives_ ? stop = true : stop = false;
  //learned_cloud += *learnable_clusters_[j];
  }

• N-samples collection in Object3dDetector::extractCluster() 直接利用human-like volumetric model分类segment

  //TODO How to classify positive/negative samples
  //===> human-like volumetric model ==> filter out over- and under-segmented clusters TODO negative!!!
  float width = max[0] - min[0];
  float depth = max[1] - min[1];
  float height = max[2] - min[2];
  if (   width  >= vfilter_min_x_ && width  <= vfilter_max_x_
  && depth  >= vfilter_min_y_ && depth  <= vfilter_max_y_
  && height >= vfilter_min_z_ && height <= vfilter_max_z_
  && min[2] <= cluster_min_z_) {
  Feature f;
  extractFeature(cluster, f);
  features_.push_back(f);
  cluster->header.seq = learnable_cluster_id_++;
  learnable_clusters_.push_back(cluster);
  } else {
  //TODO condition test only, to be removed
  //TODO N-expert based on human-like volumetric model
  if (positive_ == max_positives_ && negative_ < max_negatives_) {
      Feature f;
      extractFeature(cluster, f);
      saveFeature(f, svm_problem_.x[svm_problem_.l]);
      // -1, the negative label
      svm_problem_.y[svm_problem_.l++] = -1;
      ++negative_;
  }
  }