Doctorate thesis defense of Safa Mefteh

Doctorate thesis defense on December 13^th 2023 at 09H00 AM ,in Amphitheater Ibn Khaldoun, SUP'COM 2.

Entitled :Human Posture Recognition

Presented by :Safa Mefteh

Committee

Abstract

Human posture recognition is pivotal in various fields, including surveillance, healthcare, sports analysis, and human-computer interaction. However, automatic recognition from video sequences is challenging due to the diversity in human shapes, appearances, and complex movements. This thesis introduces a novel approach to tackle these challenges, thereby improving the performance of posture recognition. The proposed method hinges on learning local descriptors to represent human body shape and appearance.

The first part of the thesis delves into the challenges encountered by existing algorithms and offers three significant contributions to overcome them: (1) A unique multi-spectral gradient combining RGB images with depth information to identify the most relevant edges, providing crucial data about object boundaries, contours, and shape features. (2) A novel multispectral corner detector to pinpoint salient features that accurately represent the human form, based on the minimum eigenvalues of the computed multi-spectral gradient. (3) An innovative human posture descriptor calculated from the detected corner points, serving as input to the classifier.

The second part of the thesis explores classifications based on the proposed descriptor and develops a new deep model to address existing methods’ limitations. The following experiments are conducted: (1) Testing the descriptor’s effectiveness on a variety of traditional classifiers and recent models based on encoders and decoders. (2) Developing a new classifier inspired by transformers, using the descriptors to extract salient information through two different strategies. Extensive experiments on various public and widely used datasets demonstrate the proposed approach’s effectiveness, efficiency, training speed, and inference speed.

Keywords

Human Posture Recognition, Lab-D HOG Descriptor, Lab-D Corner Detector, Vision Transformer, Hybrid Learning, Computer Vision.