Wildfires pose a significant threat to the natural environment as well as public safety.Forest fire detection is critical for effective fighting, as once a wildfire has grown to a certain size, it is difficult to control. Recently, there has been a growing demand for forest areas to install a rapid response system to allow for prompt and timely action in the event of forest fires expanding across large areas. In this paper, a proposed framework for wildfire detection in a video sequence using the YOLOv5 deep learning model is presented and implemented. The interested regions represented by (fire object) in the video sequence are extracted using a new auto-annotation scheme to determine the ROI (Region of Interest) based on the edge detection process. Since the public wildfire datasets are yet confined, therefore we have constructed a new wildfire dataset named WILDFIRE-I dataset composed of variant fire images to conduct the performance evaluation of the proposed system. A comparison study with state of art research was performed in our experiments to demonstrate the efficiency of the proposed system based on common performance evaluation metrics. The experimental results exhibited detection accuracy of fire events close to 98 %, with a manual annotation process, while the proposed annotation process has achieved an accuracy of 96 %, with minimum time processing required for dataset image labelling.