Forest fires pose a severe threat to ecosystems, natural resources, climate stability, and public health. The scale and unpredictability of fires make them challenging to manage and mitigate. To minimize these impacts, an efficient fire and smoke detection system is essential. In this study, we focus on developing a camera-based fire and smoke detection system using a machine learning approach to address the inherent complexities and uncertainties in real-world fire detection. We explore various deep learning-based object detection models, including Faster R-CNN, YOLOv4, YOLOv5, EfficientDet, and DETR, leveraging soft computing’s adaptability to handle variable environmental conditions and ambiguous visual cues. Performance of each algorithm was evaluated based on mean average precision (mAP) and frames per second (fps) to balance detection accuracy with real-time processing needs. Our results reveal that DETR achieved the highest mAP at 92.2 ± 0.7%, indicating robust detection accuracy in complex scenes. However, for real-time detection requirements, YOLOv5 offered a practical advantage, achieving a high mAP of 91.0 ± 0.5% with superior processing speed.