Stream based active distillation and holonic architecture for scalable camera systems
(2024)
Files
ThèseDaniManjah_VF_8_novembre-aveccompression.pdf
Open Access - Adobe PDF
- 3.79 MB
Details
- Authors
- Supervisors
- Abstract
- Image processing is essential in camera systems for applications like traffic analysis. However, deploying video analytics solutions poses challenges such as: Continuous adaptation to the operational environment, The system's expansion with new units, and The need for context-specific solutions. This thesis aims to address these challenges by ensuring economical management of retraining and system expansion without compromising performance. Firstly, we reduce annotation and training costs per node by employing lightweight deep neural networks trained on selected samples from camera streams. We demonstrate that selecting frames with the highest confidence scores from local models produces the most accurate models. These samples are annotated using a large general-purpose model, minimizing the need for manual labeling. We also analyze potential biases introduced by model-based pseudo-annotations in comparison to human annotations. To further reduce costs, we cluster cameras based on their similarity to train group-specific models. This decreases the number of required models and enhances local performance through a specificity-diversity trade-off. Indeed, groups can fine-tune models to capture the nuances of individual streams while leveraging broader data for better generalization. We also introduce an automatic clustering strategy based on the premise that models fine-tuned on specific camera domains can effectively transfer to similar domains. Lastly, we design a holonic architecture for multi-camera and multi-method systems. By exploiting its self-contained, multi-scale, and hierarchical characteristics, we develop two key applications: (1) a multi-method framework that integrates deep neural networks with traditional image processing techniques to reduce processing costs without sacrificing quality, and (2) robust design principles that facilitate the expansion and contraction of camera systems. Through these methodologies, this thesis advances the automation and adaptability of image processing in multi-camera systems, contributing to their scalability.
- Affiliations
Citations
Manjah, D. (2024). Stream based active distillation and holonic architecture for scalable camera systems. https://hdl.handle.net/2078.5/241462