The Baseline Algorithm

Updated on 7/31/2003 11:47 AM

The baseline algorithm was designed to be simple, fast, yet effective at computing similarity of gait in video sequences, based on both shape and dynamics. The algorithm was not designed to be robust against many well know sources of variations, such as illumination, clothing, 3D viewpoint etc. Although, as we shall see, there is some amount of scale invariance built in. We have two versions of the baseline algorithm

v     the parameterized version with three parameters that need to be chosen (Version 1.*) and

v     the parameter-free version that does not require the user to select parameters (Version 2.*)


Color Code:

v     Parts unique to the Parameterized Baseline Algorithm (Version 1.*) are in GREEN

v     Parts unique to the parameter-free baseline algorithm (Version 2.*) are in RED

v     Parts common to both are in BLUE

Bounding Boxes:

Semi-automatically mark bounding boxes around the person in each frame. The user manually marks the boxes at key frames and the boxes for the intermediate frames are linearly interpolated, assuming that we have constant speed motion between the key frames. The results are output in XML format.






Gait Period Detection:

        Consider the number of silhouette pixels mostly from the legs (bottom half of the silhouettes) vs. time.

        Detect the local minima in the above plot

        Compute the median of the distances between minima, skipping every other minimum -- two possible medians, depending on whether we skipped the first one or not.

        Take the average of the medians as the gait period (Ngait).



Similarity Computation:

  • Break up probe sequence into K subsequences of Nprobe or Ngait contiguous frames each.
  • For each probe subsequence, estimate the maximum correlation with the gallery sequence.
    • Shift probe sequence with respect to gallery sequence
    • Compute distance between frame pairs
      • #pixels in AND-ed silhouettes / #pixels in the OR-ed silhouette
    • Add the distances
  • Pick the median of the maximum correlations of the probe subsequences as the similarity measure.