Performance Analysis of YOLOv5 for Real-Time Object Detection and Manipulation in Unity on CPU and GPU

Real-time object detection and interaction are essential for various applications, including human-computer interaction (HCI), augmented reality, and robotics. This study examines the efficiency of object-based interaction using an entry-level GPU (NVIDIA GTX 1050) and a CPU-only setup (Intel Core i7-9750H) with an inbuilt webcam. The pretrained YOLOv5s model was used for real-time object detection, and performance was assessed in terms of frame rate (FPS), inference speed, and responsiveness. Experimental results indicate that GPU acceleration significantly enhances performance, achieving 12 FPS compared to 8 FPS on the CPU. Without using asynchronous programming in Unity, FPS was limited to 5 on both setups, but implementing it within the Unity’s C# script increased FPS to 12 on the GPU and 8 on the CPU. This demonstrates that software optimizations can improve real-time performance even on resource-limited hardware. A comparative analysis with prior studies highlights the trade-offs between different YOLO versions and hardware configurations. Our findings suggest that real-time object interaction can be efficiently achieved on entry-level hardware, making cost-effective HCI solutions feasible. This research contributes to the development of interactive applications by optimizing computational resources without requiring high-end systems.