15 Daniele Berardini Computer Vision News My thesis aims to contribute to the integration of Edge AI in Computer Vision by focusing on the design and development of lightweight, deep learning-based monitoring systems for the real-time analysis of images and videos. It explores two domains, each with its own challenges: security, focusing on weapon detection in surveillance videos, and healthcare, focusing on segmenting preterm infants’ limb poses from depth data. Regarding the security domain, in collaboration with INIM Electronics, an Italian leader in security systems, I tackled weapon detection challenges, notably the difficulty in identifying small-sized weapons and the need for real-time weapon localization. Current solutions, like image Super Resolution (SR) methods or complex detection architectures, are inapplicable on edge devices due to computational constraints. To address this, after the creation of a surveillance dataset for weapon detection (WeaponSense), I proposed the first Edge AI framework for real-time weapon detection in surveillance videos through the use of two cascaded CNNs, optimized for edge devices. Despite the results improving the state of the art, the framework shows limitations on efficiency in crowded environments. Thus, in collaboration with the University of Córdoba (Spain), I proposed a novel method that integrates during training an Enhanced Deep Super Resolution network into an edge-oriented CNN for weapon detection, discarding the former during inference. The proposed approach overcomes the previous limitations, enabling accurate and real-time on-device localization of weapons (Figure 1). In the medical domain, my research was motivated by the need for automated technology to continuously monitor the movement of preterm infants, which is essential for early assessment of potential long-term complications. Current methods are effective but require very high operational costs, hindering their implementation in budgetlimited facilities. Driven by these premises, I initially proposed a CNN that incorporates lightweight computational blocks from a segmentation network (EDANet) into the bi-branch structure of a preterm infants’ pose segmentation network (BabyPoseNet). Subsequently, by conducting a per-layer complexity analysis of the proposed CNN, I redistributed computation throughout the network to reduce complexity. This approach yielded a real-time framework capable of running on edge devices, achieving optimal accuracy in segmenting infants’ limbs. Ultimately, my thesis aims to push research toward more sustainable and affordable AI solutions in different domains, valuing not only high accuracy, but also the efficiency and adaptability across various needs and contexts.
RkJQdWJsaXNoZXIy NTc3NzU=