Minimizing damage and improving safety in public and private areas depend on timely and precise fire detection in video surveillance. High false favorable rates and slow reaction times are common problems with traditional fire detection techniques, especially in visually complex environments. An improved YOLOv5 architecture designed for effective fire recognition is proposed in this study, which overcomes these drawbacks by increasing detection accuracy and speed. To capture the distinct qualities of fire, such as irregular shapes, fluctuating movement, and varying intensities, key modifications include incorporating an improved feature extraction layer and optimised anchor box parameters. In real-time surveillance applications, experimental findings indicate that the enhanced YOLO v5 model demonstrates superior performance compared to traditional fire detection approaches, achieving higher precision and recall rates with minimal computational overhead. Compared to YOLOv5s and YOLOv5x, the proposed model achieves an improvement in mAP of 0.03 and 0.01 compared to YOLOv5s and YOLOv5x respectively, when evaluated on a CPU thereby demonstrating enhanced performance and speeds up inference by up to 5.47 and 21.31 FPS. As a result, our model can be used for effective fire detection in real-world scenarios.