Inference and Performance Analysis of Optimized YOLOv9 for Indoor Object Detection & Classification

Indoor object detection presents unique challenges such as occlusions, varying lighting conditions, and cluttered environments. While several object detection frameworks, including RetinaNet, Faster R-CNN, SSD, and EfficientDet, have been proposed, they often suffer from high computational cost, reduced inference speed, and limited accuracy in terms of mean Average Precision (mAP), particularly in real-time scenarios. In this study, lightweight YOLO variants, namely YOLOv7, YOLOv8s, YOLOv9s, and a fine-tuned YOLOv9s which considers the optimized training strategy based on albumentations. All the models are evaluated for indoor object detection using the RGB TUT Indoor dataset. The models are assessed using precision, recall, mAP@0.5, and mAP@0.5:0.95. The experimental results demonstrate that the fine-tuned YOLOv9s consistently outperforms the baseline YOLOv9s model across all evaluation metrics, confirming the effectiveness of proposed training optimizations. Specifically, the fine-tuned YOLOv9s achieves a precision of 97.9%, recall of 96.1%, mAP@0.5 of 99.1%, and mAP@0.5:0.95 of 88.7%. These improvements highlight the impact of systematic training refinement beyond standard model configuration. Among the evaluated models, YOLOv8s achieves the highest inference speed of 90 FPS in 11.1 ms, making it suitable for ultra-low-latency applications such as smart homes, assistive systems, and robotics. In contrast, the fine-tuned YOLOv9s provides a superior balance between accuracy and efficiency, making it more suitable for accuracy-sensitive indoor environments where detection reliability is critical. Overall, the study demonstrates that carefully optimized training strategies can significantly enhance the performance of YOLOv9s without architectural modifications, providing practical insights for real-time indoor object detection systems.