Thesis
PhD Thesis
Title: Person Identification Through Footstep-Based Biometric Recognition
Advisor: Prof. Subrat Kar
Description:
During my PhD, I am working on footstep-based person identification problems using a seismic sensor
A person verification system (analogous to a speaker verification system) based on footfall signatures is proposed using Gaussian Mixture Model - Universal Background Model(GMM-UBM). The robustness of the system is evaluated by creating different scenarios of registered and non-registered users. A performance comparison with the prior art demonstrated the superiority of the proposed system.
A novel unconstrained biometric authentication utilizing footstep-induced seismic signal is proposed. An Unsupervised Learning based Event Extraction Technique(USLEET) is proposed for extracting the footstep events from the raw signal. Following that temporal and spectral features are extracted from these footstep events. To detect unauthorised users(imposters), one-class classifiers such as One-Class Support Vector Machine(OC-SVM) and Support Vector Data Descriptor(SVDD) are utilized. The real-world scenarios are mimicked by training these classifiers by using only the footstep data from the registered users. The footsteps of the unregistered users are unseen by these trained classification models. The performance is evaluated on the dataset containing ~78000 footsteps from eight individuals.
A Robust person identification system by using the supervised machine learning algorithm is proposed for smart home applications. A three-layer computing architecture is designed to increase the scalability of the system and minimize the latency. It also eliminates the need to store the raw data in the central repository by processing the raw signal at the edge of the network. A basis pursuit-based footfall event data compression technique is also proposed for the effective transmission of the raw signal to the edge of the network. The proposed framework is evaluated using the indigenous dataset of over 100,000+ footfall.
A convolutional neural network(CNN) with transfer learning called FootsNet is proposed for person identification. It addressed the problem of identifying new users in the system. The existing state-of-the-art techniques cannot accommodate the new users and require fresh training in the person identification model. The retraining of the model requires a large amount of labelled data which is, again, very cumbersome and time-consuming work. The proposed methods solve this problem by adapting the two-stage training process. First, the identification model is pre-trained with the base classes(existing users), which have a large amount of labelled data. Then, the model is fined-tuned using the limited data from the novel classes(new users).
MS [Res.] Thesis
Title: Human Identity Characterization and Classification using a Seismic Sensor
Advisor: Prof. Subrat Kar
Description:
The walking human subjects are classified from their footsteps using a seismic sensor
Novel pre-processing event detection techniques are proposed to extract signals of interest from the raw seismic signal
Temporal and spectral-domain features are extracted from footstep events, which help determine the distinct characteristics of each class
The extracted features are subsequently utilized to train a supervised machine-learning classifier model. The trained model is then employed to classify the testing input features