Chapter 2 Literature Review
2.2 Gait Recognition Challenges and Databases
In this section, we introduce gait recognition challenges and the commonly used gait databas- es consisting of images with these challenging factors for algorithm evaluation.
2.2.1 Challenges
In real-world scenarios, the walking conditions between the probe and gallery images are not always the same. Walking conditions may change due to different covariate factors, which can be divided into four categories:
1. speed,
2. camera viewpoint,
3. covariates with unpredictable effect, e.g., shoe type, carrying condition, clothing, walking surface, elapsed time, etc.
(a) (b) (c) (d) (e) (f) (g)
Figure 2.2: GEIs of one subject walking in different walking conditions from the USF gait dataset [Sarkar et al., 2005]. (a) is the GEI in normal condition. (b)-(g) are the GEIs under the influences of (b) viewpoint, (c) walking surface, (d) viewpoint and walking surface, (e) carrying condition, (f) carrying condition and viewpoint, (g) elapsed time, shoe type, clothing, and walking surface.
Figure 2.3: GEI examples from CASIA-B dataset [Yu et al., 2006] of a subject from view
0◦to180◦, with an interval of18◦.
Since small changes in speed and camera viewpoint only have limited impact on gait recog- nition ([Tan et al., 2006], [Yu et al., 2006]), it is more reasonable to evaluate several rep- resentative speeds/views, instead of all the possible speeds/views. In OU-ISIR-A dataset [Makihara et al., 2012], the walking speeds for evaluation are from 2km/h to 7km/h, with an interval of 1km/h, while in CASIA-B dataset [Yu et al., 2006], the camera viewpoints for evaluation are from0◦to180◦, with an interval of18◦. For detailed studies on the effect of speed/viewpoint, we separate these two (relatively) controllable covariates from others.
Despite the effectiveness of GEIs, when the walking condition changes, matching GEIs directly makes the classification prone to errors. Fig. 2.2 shows some GEIs of the same subject walking in different walking conditions from the USF gait dataset [Sarkar et al., 2005], while Fig. 2.3 demonstrates several GEIs of the same subject in different views from the CASIA-B dataset [Yu et al., 2006]. We can see that covariates may substantially alter the human appearance, thus giving rise to recognition difficulties.
Figure 2.4: Gait images with different speeds from the OU-ISIR-A dataset[Makihara et al., 2012]
2.2.2 Databases
OU-ISIR-A
The OU-ISIR-A dataset [Makihara et al., 2012] was constructed for evaluating speed- invariant gait recognition algorithms. It was collected on a treadmill with a large range of speeds (from 2km/h to10km/h) in terms of walking or running for 34 subjects. The sub- jects were instructed to walk at six different speeds, ranging from 2km/h to 7km/h with an interval of 1km/h, and to run at three different speeds, ranging from 8km/h to 10km/h with an interval of 1km/h . Several gait images from this dataset are shown in Fig. 2.4.
OU-ISIR-B
The OU-ISIR-B dataset was constructed by Hossain et al. [Hossain et al., 2010] for studying the effect of clothing on gait recognition. It includes 68 subjects walking on a treadmill with up to 32 types of clothes combinations. Several gait images from this dataset are shown in Fig. 2.5.
OU-ISIR-D
The OU-ISIR-D database [Makihara et al., 2012] consists of two datasets, namely, DB-high (i.e., with small gait fluctuations) and DB-low (i.e., with large gait fluctuations). For DB- high/DB-low, there are 100 subjects (1 subject per sequence) for both the gallery and probe. The original resolution and frame-rate in OU-ISIR-D database are128×88pixels and 60
Figure 2.5: Gait images with several different clothes types from the OU-ISIR-B dataset [Hossain et al., 2010]
Figure 2.6: Gait images from CASIA-B dataset [Yu et al., 2006] of a subject from view0◦
to180◦, with an interval of18◦;
fps. This dataset is normally down-sampled in a spatial and temporal manner, and used to test gait recognition algorithms on extremely low quality videos.
CASIA-B
The CASIA-B gait dataset [Yu et al., 2006] is a large multi-view gait dataset, which consists of 124 subjects walking in the indoor environment with the cameras fixed at 11 viewpoints (from0◦to180◦with an interval of18◦), as shown in Fig. 2.6.
Figure 2.7: Gait images from CASIA-C dataset [Tan et al., 2006] of a subject collected at night environment using infrared cameras two different walking conditions: normal walking and carrying condition, from left to right
CASIA-C
The CASIA-C dataset [Tan et al., 2006] was collected at night time using infrared cameras, with 153 subjects in three different speeds (i.e., slow/normal/fast walking) and a carrying condition. Two gait images from this dataset are shown in Fig. 2.7.
TUM-GAID
This TUM-GAID dataset [Hofmann et al., 2014] simultaneously contains RGB images, depth images, and audio of 305 subjects in total. In [Hofmann et al., 2014], Hofmann et al. designed an experimental protocol (based on 155 subjects) to evaluate the robustness of algorithms against covariate factors like shoe, carrying condition (5kg backpack), elapsed time (January/April) which also potentially includes changes in clothing, lighting condition, etc. Several gait images from this dataset are shown in Fig. 2.8.
USF
The USF dataset [Sarkar et al., 2005] is a large outdoor gait database consisting of 122 subjects. A number of covariate factors are considered: camera viewpoints, shoes, surface types, carrying conditions, elapsed time, and clothing. Several gait images from this dataset are shown in Fig. 2.9.
Figure 2.8: Gait images from TUM-GAID dataset [Hofmann et al., 2014] of a subject with six different walking conditions: normal walking (in Jan.), with a backpack (in Jan.), with a different pair of shoes (in Jan.), normal walking (in April), with a backpack (in April), with a different pair of shoes (in April), from left to right
Figure 2.9: Gait images in the outdoor environment from the USF dataset [Sarkar et al., 2005]