Clothing variations bring many difficulties to gait recognition. We proposed a part-based approach to conquer the problems brought by clothing variations. We partition a human body into several parts, including overlapping parts. The overlapping is important because clothing variations can be approximated by combinations of different parts. The key ideas in this paper are listed as follows.
Part-based clothing categorization
We established a clothing categorization to group similar clothes. In the training phase, Probability Distribution Functions (PDFs) between gait features are stored separately for each part and for both the same and different clothes, and the corresponding discrimination capabilities are also measured in advance. In the test phase, given a probe, posteriors with the same and different clothes, are calculated based on the distances between the probe and the galleries and on the trained PDFs.
Adaptive weight control
In a part-based method, the main issue is how to combine the individual distances into a single distance that quantifies the overall matching measure between a probe and the gallery. We adopt the weighted sum of distances corresponding to each part as the overall matching measure and adaptively control the weights for better identification, by increasing them for the same part of clothing and vice versa.
We define 4 alternative options for the matching measure. The first option is the mean value of the minimum distances of the subsequences for each probe and gallery. The second is the median value of the minimum distances of the subsequences for each probe and gallery. The third is the minimum value of the minimum distances of the subsequences for each probe and gallery. We propose a new technique using the average of the 2 minimum distances of the subsequences for each probe and gallery. The experimental results show that proposed measure gives the best results. We have therefore chosen the average of the 2 minimum values, as our matching measure.