Mohammed Guermal

Building a computer vision system for real-time drowning detection in pools, from swimmer tracking to alert generation Explored and benchmarked detection/tracking pipelines for swimmers: YOLO, SAM3, SORT, DeepSORT, Ultralytics -- evaluating trade offs between tracking robustness and real-time constraints in a challenging visual environment (water, occlusion, reflections) Designed video classification models for real-time drowning-alert detection using 3D CNNs and Transformer architectures Led the creation of a purpose-built swimmer-tracking dataset, since no public dataset covers pool-based drowning scenarios Researching how vision-language models (Qwen, LAVAD, LLaVA-NeXT) can be adapted as a downstream reasoning layer on top of raw detections, to add contextual interpretation to drowning alerts Applying temporal action-anticipation techniques (LSTR, TesTra) -- extending the online action-anticipation research from my PhD (see JOADAA, WACV 2023) -- to predict drowning risk before it fully develops, rather than only detecting it after the fact

Built neural network models for theft detection in retail video streams, combining 3D CNNs and Transformer-based architectures for anomaly/action recognition -- an industry application closely related to my anomaly-forecasting research (see "Guess Future Anomalies from Normalcy", WACV 2025) Designed multimodal models fusing skeleton, RGB, and text modalities, extending my PhD work on multimodal fusion (see CM3T, WACV 2025) to capture complex movements and video context beyond a single modality Explored self-supervised pretraining to reduce dependence on annotated data: experimented with SimCLR, MAE, VideoMAE, and V-JEPA to learn general-purpose video representations before fine-tuning on theft-detection tasks Investigated contrastive learning for video-text alignment -- SigLIP, CLIP, X-CLIP -- to identify suspicious behaviors and thefts across long, weakly-labeled video sequences Applied knowledge distillation to compress models for embedded/edge deployment on in-store storage servers, reducing memory footprint and latency (up to 60% in production tests) without materially sacrificing detection accuracy Integrated large vision-language models (Qwen via Hugging Face, LLaVA-NeXT Video) as a contextual-reasoning layer on top of detection outputs, to interpret ambiguous actions rather than relying on classification alone Ran full training pipelines on AWS GPU infrastructure, from data/resource management to deploying prototypes for real-world validation

Developed and optimized human pose estimation algorithms to detect and track human movements in video, targeting applications in spatial/aerospace environments Designed unsupervised domain adaptation methods to transfer object detection models to challenging, unfamiliar visual environments without annotated data -- addressing the common gap between lab-trained models and deployment conditions Investigated video-based pose estimation across varying lighting and visibility conditions to improve robustness beyond standard benchmark settings Proposed new algorithmic approaches for pose estimation under limited-data settings, where annotation is expensive or unavailable Collaborated across disciplines with AI researchers and aerospace industry experts to integrate research prototypes into real-world applications