Authors - Manisha Mane, Saurav Bedse, Vikrant Patil, Pruthviraj Dhande, Om Darekar Abstract - The fast advances in deep learning and computer vision have dramatically improved the ability to detect objects, with applications in surveillance, driverless cars, and smart traffic management. The current paper describes an implementation of the YOLOv8 model for real-time object detection on different categories such as persons, cars, and bicycles. We trained the model on a customized dataset of annotated images, fine-tuning it through extensive hyperparameter tuning and multiple training epochs. Our training setup consisted of 75 epochs, utilizing a Tesla T4 GPU for computation. The model recorded a mean Average Precision (mAP@50) of 76.5% over all classes, with class performance highlighting high precision and recall rates for classes like cars (98.2%) and bicycles (87.8%). To further improve accuracy, we utilized data augmentation methods, batch normalization, and optimizer tuning. After training, the model was subjected to extensive validation, with an inference speed of 8.5ms per image, making it viable for real-time performance. We also incorporated the model into a realistic deployment pipeline, showcasing its efficacy in real-world applications. This paper presents a thorough analysis of the trained model, such as performance metrics, comparison with other versions of YOLO, and discussion of future improvements. Our results emphasize the model’s ability to achieve speed and accuracy balance, rendering it an appropriate choice for object detection in real-time applications. Future research will investigate additional optimizations such as light-weight model variants and domain-specific dataset adaptation.
