Human Gait Recognition Under Changes of Walking Conditions
- Publication Type:
- Thesis
- Issue Date:
- 2021
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Gait has been gathering extensive research interest for its non-fungible position in applications. First, it is difficult to disguise one's gait, since walking is necessary for human mobility. Second, it can be attained at a distance without physical contact or proximal sensing. However, although recently different methods have been proposed for gait recognition, gait analysis is still in its infancy. Most methods enable to garner a remarkable recognition performance when the gallery and the probe are in a similar situation. However, when exterior factors affect a person's gait and changes occur in human appearances, a significant performance degradation happens.
Among these exterior factors, clothing variations and mode changes can be treated as the most influential factors for gait recognition. Clothing variations can significantly influence available features to be used in the future recognition process, while walking/running modes can change human motions made by limbs and thus dramatically influence the instinct walking patterns of each person. Thus, in this thesis different methods have been proposed for gait recognition to handle the difficulties of clothing variations and walking/running mode changes.
First, given that model-based methods are less vulnerable to clothing variances, a more robust model-based gait feature, Skeleton Gait Energy Image (SGEI), is formed to handle this cloth-changing gait recognition problem. Then, since clothing changes can cause different impacts to different body parts, a part-based collaborative spatio-temporal feature learning method is also proposed for cloth-changing gait recognition by concatenating features from the non/less affected body parts under the correlative H-W and T-W views. Based on the aforementioned two methods, another efficient network is proposed for cloth-changing gait recognition. This network consists of two sub-networks, aiming to produce part-based features from the non/less affected body parts and the estimated skeleton key-point regions. Moreover, in order to address the walking-vs-running problem in a cross-mode manner, a feasible hybrid method is also proposed in this thesis. Distinct from most cross-mode gait recognition methods, this method focuses on learning mode-invariant features for each person from their innate patterns between walking and running modes. Multi-task learning strategies are also used to enhance the efficiency of these learned features. Finally, given that the above-mentioned methods are all proposed based on sufficient input data, a complementary solution is given when only a few gait frames can be offered.
Experiments have proved that these proposed methods can obtain a remarkable performance when tackling the cloth-changing and walking-vs-running gait recognition problems.
Please use this identifier to cite or link to this item: