Maximizing YOLOv2 efficiency: A study on multiclass detection of indoor objects

• Optimization of YOLOv2 hyperparameters for high accuracy automated detection of seven types of indoor objects. • Base CNN model compared among Tiny and DarkNet-19 networks. • Demonstrates the importance of hyperparameter optimization in multiclass detection with substantial class and object size imbalances. The objective of the present study is to present a systematic approach for optimizing the key hyperparameters of YOLOv2 model for multiclass object detection, specifically targeting seven classes of indoor objects: chair, fire extinguisher, printer, screen, trash bin, exit, and clock. Based on an indoor object detection dataset containing 2213 images covering varying lighting conditions, backgrounds, occlusion and high inter-class differences, the YOLOv2 network is refined to enhance detection performance by examining the impact of different deep CNN architectures (Tiny-Coco, DarkNet-19-Coco), solvers (SGDM, RMSProp, Adam), learning rates (10 −5 , 10 −4 , 10 −3 ), and detection thresholds (0.1, 0.2, 0.3) on mean average precision (mAP) and detection scores. The findings indicate that both Tiny-Coco and DarkNet-19-Coco models achieve optimal detection performance with a learning rate of 10 −5 , a threshold of 0.1, and the Adam optimizer. Notably, lower thresholds (0.1) and slower learning rates (10 −5 ) consistently yielded superior results, with 5-10% and 20-30% higher detection performance metrics compared to their higher levels. The Adam solver was identified to be the most effective, significantly enhancing model performance, with around 5-10% higher performance metrics compared to other solvers. Deviations from these optimal settings resulted in decreased accuracy and reliability. Therefore, it is crucial for object detection algorithms to incorporate optimized parameters to maximize efficiency, particularly in multiclass detection scenarios with substantial class and object size imbalances, as demonstrated in this work. These findings have significant implications for real-world applications, such as smart home systems, office automation, and surveillance, where efficient and accurate indoor object detection is critical.