Siamese Neural Network based Gait Recognition for Human Identification Cheng Zhang, Wu Liu, Huadong Ma, Huiyuan Fu Beijing Key Laboratory of Intelligent Telecommunication Software and Multimedia, IEEE International Conference Beijing University of Posts and Telecommunications, Beijing, China, 100876 on Acoustics, Speech and {zhangcheng, liuwu, mhd, fhy}@bupt.edu.cn Signal Processing
1. Introduction 3. Siamese Neural Network based Gait Recognition 4. Experiments
As the remarkable characteristics of remote accessed, OULP-C1V1-A Gait Database robust and security, gait recognition has gained signi- - Contains the world’s largest number of subjects (3835). ficant attention in the biometrics based human identi- fication task. - Records two sequences for each subject: probe (query) and gallery (source) sequence with 4 observation angles. The existed methods mainly employ the hand-crafted gait features, which cannot well handle the indistinctive Pipeline inter-class differences and large intra-class variations of - Foreground segmentation → Periodic identification → human gait in real-world situation. GEIs generation → SiaNet training using gallery set → To this end, we present one of the first attempts to study Feature extraction → K-Nearest-Neighbor searching the Siamese neural network (SiaNet) based gait recogni- Evaluation on Intra-view Human Identification tion framework for human identification with distance metric learning. Fig.1 The framework of our proposed gait recognition method.
Different from conventional CNN, the Siamese network Solutions can employ distance metric learning to drive the simi- - To solve the data limitation problem, we use GEI instead of raw sequence of hu- larity metric to be small for pairs of gait from the same man gait, which can remove most noisy information while keep the major human Table.1 Comparison results of different methods in terms of the person, and large for pairs from different persons. shapes and body changes during walking. Rank-1 and Rank-5 Identification Rates In the end-to-end framework, we leverage the competiti- - To close the domain gap between recognition and classification, we adopt the Evaluation on Inter-view Human Identification ve Gait Energy Image (GEI) presentation as the input of Siamese neural network, which can simultaneously minimize the distance bet- 0.9 SiaNet AVTM_PdVS AVTM woVTM RankSVM network while holistically exploit the Siamese neural net- 0.85 ween similar subjects and maximize the distance between dissimilar pairs with a work to learn effective feature representations for human 0.8 distance metric learning architecture. identification. 0.75 The Structure of the Siamese Neural Network 0.7
- The Siamese neural network designed for gait recognition contains two parallel 0.65 2. Conventional CNN based Gait Recognition 0.6 CNN architectures sharing the same parameters. 0.55 Approach - The input of the network are pairwise GEIs with similar or dissimilar labels. 0.5 ( 65°A, 75°) ( 75°,B 65°) ( 75°C, 85°) ( 85°D, 75°) - We attempt to fine-tune the conventional CNN on the - The output of the CNNs are combined by the contrastive layer to compute the Fig.2 Comparison of the cross-view matching approaches on
gait database for gait recognition based human identi- contrastive loss. different types of inter-view test in terms of Rank-1 Id-Rate.
fication task. Differently, we change the output of CNN to Training and Learning the number of subjects in the gait database. - The distance E W (,) x 12 x between a pair of GEIs can be measured by: 5. Conclusions and Future Works Problems 2 EWWW(,)()() x1 x 2 S x 1 S x 2 We have developed a Siamese neural network based - Data limitation. CNN requires a mass of training data 2 - We can define the contrastive loss function as follows: gait recognition framework for human identification with for all categories. For gait recognition, the number of su- P i distance metric learning, which outperforms the state-of- bjects can be large, while with only a few examples per ()(,(,,))W L W y x12 x i1 the-arts on the world’s largest challenge gait database. subject in public database. LWyxx( ,( , , )i ) (1 y ) max( mExx ( , ) i ,0) yExx ( , ) i 1 2WW 1 2 1 2 In the future, we will try to train 3-Dimensional Siamese - Domain gap. Gait recognition for human identification is - The contrastive loss function can be minimized over a set of training pairs using neural network with more training dataset to further essentially a search problem but not classification. stochastic gradient descent. improve the performance of the gait recognition.