运动目标图像序列聚类异常：DBSCAN单帧聚类多簇问题求解

Great question—this is a common pain point when using DBSCAN on segmented pedestrian imagery, especially if background removal leaves tiny disconnected regions or random noise pixels. Here are actionable ways to leverage prior knowledge to eliminate those unwanted clusters and get the single pedestrian cluster you expect:

• Tune DBSCAN parameters using pedestrian size priors
  DBSCAN’s eps (maximum distance between two points to be in the same neighborhood) and min_samples (minimum points required to form a core point) make all the difference here. Since you know the approximate size of a pedestrian in your frames (e.g., average height/width in pixels), use this to set these values strategically:

  • Set eps to a value slightly larger than the maximum gap you’d expect within a single pedestrian (e.g., if your mask has small gaps up to 3 pixels wide, set eps=4).
  • Set min_samples to the minimum number of pixels a valid pedestrian should have (e.g., if the smallest pedestrian in your dataset is 200 pixels, set min_samples=150 to filter out tiny noise blobs).
    This ensures disconnected parts of the same pedestrian are grouped together, while random noise (usually just a few pixels) gets labeled as outliers.

• Post-filter clusters using pedestrian shape/aspect ratio priors
  Even after tuning DBSCAN, some spurious clusters might slip through. Use your knowledge of typical pedestrian dimensions to weed them out:

  1. For each cluster, compute its bounding box.
  2. Calculate the aspect ratio (height/width) and area of the bounding box.
  3. Discard any cluster where:
     • The area is below a threshold (e.g., less than 100 pixels—way smaller than a real pedestrian).
     • The aspect ratio is outside the typical range for pedestrians (usually between 2:1 and 3:1; a cluster with a 1:5 ratio is almost certainly noise).
       You can also merge adjacent clusters if their combined bounding box fits the pedestrian aspect ratio—this handles cases where DBSCAN split a single pedestrian into two small clusters.

• Leverage temporal consistency across frames
  Since you’re working with a sequence of 100 frames, pedestrians move smoothly over time, while noise clusters are random and won’t persist across frames. Use this temporal prior to clean up results:

  • For each frame, track the position and size of valid pedestrian clusters from the previous frame.
  • If a small cluster in the current frame doesn’t have a corresponding cluster in the previous/next frame (or its position change is way faster than a typical pedestrian’s walking speed), mark it as noise.
    This is especially effective because random noise clusters are one-off and won’t have consistent motion.

• Pre-process masks to fill gaps before DBSCAN
  Often, spurious clusters come from tiny gaps in the pedestrian mask (left by imperfect background removal). Use morphological operations to fix this before running DBSCAN:

  • Apply a closing operation (dilation followed by erosion) to fill small holes and connect disconnected regions in the mask. The kernel size for dilation/erosion can be based on your prior knowledge of how big gaps in the pedestrian mask are (e.g., a 3x3 kernel for small gaps).
  • This turns the pedestrian into a single connected blob, so DBSCAN will cluster it as one instead of splitting into multiple parts.

• Modify DBSCAN to enforce constraints (advanced)
  If you want to go a step further, you can tweak the DBSCAN algorithm to incorporate pedestrian shape priors during clustering. For example, when checking if a point should be added to a cluster, verify that the growing cluster still fits the expected aspect ratio and size bounds. This prevents small noise regions from forming their own clusters in the first place, though it requires more code changes than post-processing.

All these methods rely on the prior information you have (pedestrian size, shape, motion) to distinguish valid clusters from noise—exactly what you need to fix your issue.

内容的提问来源于stack exchange，提问作者Manish Sharma