Abstract: Transmission lines are often built in sparsely populated areas such as mountains and forests. If a wildfire occurs near this area, it will seriously threaten the social economy and the safety of the people. Fixed cameras are usually used to monitor early wildfires in real time. Early wildfires often appear in the form of smoke. Based on the characteristics of fixed cameras, we propose a pixel-oriented feature extraction method. Gaussian filtering was used to further process the features, and abnormal points in the difference matrix were detected by anomaly detection method. Finally, the fire point was marked and displayed. The experimental results show that compared with the supervised classic machine learning method, the method based on anomaly detection does not need to be trained in advance, and can accurately detect very small smoke with the contextual information advantages of fixed camera shooting. And it can increase samples of early wildfire images and balance positive and negative samples. Moreover, the algorithm has a simple structure and is suitable for embedded cameras to complete front-end intelligent analysis and achieve purpose of real-time monitoring.