Abstract
Object detection plays a critical role in automating visual recognition tasks like fall detection, which minimizes the need for constant human intervention. However, environmental challenges such as low-resolution images, crowded spaces, and varying image scales can decrease detection accuracy and increase the false positive rates. The proposed work aims to develop an adaptive, ensemble-based fall detection model named "YORD," which integrates modified versions of YOLO V8, RetinaNet, and baseline DETR. The goal is to improve detection accuracy and robustness across diverse environmental conditions. The YORD model leverages YOLO V8 for its speed and accuracy in real-time detection, DETR's attention mechanism for capturing falls from multiple angles and low-resolution images, and RetinaNet's focal loss strategy to address class imbalance and improve multi-scale detection. The dataset is enhanced using Albumentation to simulate real-world fall scenarios. The model is implemented in Google Colabs for streamlined experimentation and analysis. Experimental findings show that the YORD model is highly reliable and robust, with reduced false positives by utilizing the complementary strengths of each component model through ensemble learning. The YORD model achieved precise fall detection results across various challenging conditions, explaining the enhanced performance over the individual models. The YORD ensemble model effectively addresses the limitations of single-model fall detection systems, offering a robust solution for real-time applications. However, the system still faces some limitations in identifying the falls in the occluded images where the objects are only visible partially making it harder for the system to localize and classify them correctly. Future work will focus on fine tuning certain aspects of the model to focus on important parts of an image and give less importance to irrelevant or occluded areas.