Motivation Description
To protect and preserve environments in surface, air, underwater and space using active or passive sensors. (Short Presentation)
General Description
The conservation of the environment is a matter of global importance, requiring the implementation of advanced technologies for intelligent monitoring and protection of natural ecosystems. In practice, two principal ways are employed to preserve the environment:
The first one is the observation of human activities in order to regulate and limit their impact. This could involve optimizing road traffic, minimizing industrial emissions or detecting behavioral anomalies. In this context, passive monitoring exists since the emergence of cameras. The moving objects are terrestrial entities (humans, vehicles), marine entities (boats) and aerial entities (airplanes, drones). (More information)
The second manner is to carry out the monitoring of the state of fauna and flora and its evolution collecting various data could help to understand the dynamics of ecosystems, detect changes, and develop preservation strategies. In this context, passive monitoring has emerged as a transformative tool for applied ecology, conservation and biodiversity monitoring. The moving objects are terrestrial animals (mammals, birds, insects, etc) and marine animals (marine mammals, fish, ect). (More information)
Nowadays, these tasks are usually performed by means of various computer vision techniques for vision monitoring, and various sound detection and classification techniques for acoustic monitoring. For both, the acquisition of data is made via various passive and/or active sensors. These application fields are called Active and Passive Vision Monitoring (AVM/PVM) and Active and Passive Acoustics Monitoring (AAM/PAM), respectively. My research find applications in these domains. (Remarks).
These domains provide data coming from terrestrial and underwater environments using:
Passive visual sensors (visible spectrum, IR cameras, multi-spectral cameras, hyper-spectral cameras, light-field cameras and event cameras) and active visual sensors (SAR, PolSAR, and LiDAR) providing RGB data, RGB-D data, IR data, multi-spectral data, hyperspectral data, etc.
Passive acoustics sensors (microphone, microphone arrays, hydrophone, bioacoustic sensors) and active acoustic sensors (acoustic Doppler sensors, Distributed Acoustic Sensing) providing signal 1D data, etc.
The abundant data produced by these diverse sensors in both video surveillance and acoustic monitoring are impossible to process and analyze by human beings, and this requires intelligent systems based on advanced processing technique.
Data Description
Acoustic and visual sensors provide an extremely diverse and large amount of data with their own characteristics. (More information)
The challenges met in the data varies in terms of:
Imprecision, uncertainty and imcompletnesss.
Volume. [Big Data]
Dimension [Mathematics]: Scalars, vectors, matrices, and tensors. [Multi-dimensional Data]
Dimension [Signal Processing] :1-D, 2-D, 3-D, etc. [Multi-dimensional Data]
Distribution: Homogeneity and heterogeneity. [Various Sensors]
Distribution: Homoscedasticity and heteroscedasticity (i.e not the same variance).
Noise distribution (Gaussian, Laplacian, etc).
They can be handled with mathematical concepts, machine learning methods, and signal processing techniques.
Theory Description
Mathematical concepts: Crisp concepts, Statistical concepts, Fuzzy Concepts (Zadeh 1965), Dempster-Schafer (Dempster 1968, Schafer 1976). Statistic, fuzzy, and Dempster-Shäfer methods take into account imprecision, uncertainty, and incompleteness in the data, which appear due to different challenges, and because of the quality of the sensors.
Machine Learning Concepts:
Representation Learning (Dimensionality Reduction, Subspace Learning): Conventional Subspace Learning (Pearson 1901), Robust Subspace Learning (Candès et al. 2009), Dynamic Subspace Learning (Vaswani et al. 2010). Data are considered as inliers data, outliers data or missing data.
Neural Networks Learning: Conventional Neural Networks (Rosenblatt 1957) , Support Vector Machines (Vapnik et al. 1995), Deep Neural Networks (Le Cun). Data are viewed as learning entities that can be well labeled or noisy labeled.
Robust representation learning models (model-based techniques, unsupervised techniques) seek for low-rank structure in the data and attempt to be robust to outliers whereas deep Learning models (data-based techniques, supervised techniques) seek to be able to face all the challenges by learning all the situations.
Constrative learning allows to learn embedding space in which similar sample pairs stay close to each other while dissimilar ones are far apart. Contrastive learning can be applied to both supervised and unsupervised settings. In the context of unsupervised data, contrastive learning is one of the most powerful approaches in self-supervised learning.
Graph learning is a branch of machine learning that focuses on the analysis and interpretation of data represented in graph form. A graph is a collection of nodes and edges, where nodes represent entities and edges represent the relationships or interactions between these entities. Graph learning is the most powerful approach in semi-supervised learning.
Interpretability of machine learning concepts can be done with algebra (Vidal et al., 2017) and geometry (Lei et al, 2020).
Signal Processing Concepts: Estimation Filters (Kalman 1960), Sparse Representation, Graph Signal Processing (Ortega et al. 2017). Data are viewed as signal data (1D, 2D, or 3D).
Classification/Clustering Concepts: Crisp concepts, Statistical concepts, Fuzzy concepts (Bezdek 1978). Data are viewed as belonging or not to a class or a cluster.
Implementations Description
The algorityhms developed for AVM/PVM and AAM/PAM needs to be implemented on computers or embedded systems achieving real-time performance:
CPU Implementations: A central processing unit (CPU) is the primary processor in a given computer. The language C and the library Open CV allow to implement algorithms on CPU.
GPU Implementations: A graphics processing unit (GPU) is a specialized electronic circuit designed for digital image processing and to accelerate computer graphics, being present either as a discrete video card or embedded on motherboards. The language CUDA allows to implement algorithms on GPU to achieve parallel implementations.
FPGA Implementations: Field Programmable Gate Arrays (FPGAs) are integrated circuits often sold off-the-shelf. They referred to as field programmable because the customers can reconfigure the hardware to meet specific use case requirements after the manufacturing process. The language VHDL allows to program FPGA to achieve real-time performance.