In order to solve the problem of leakage detection in the case of scarce visible foggy gas leakage sample data and lack of targeted features, a multi-feature based visual detection method for foggy gas leakage is investigated. Using the differences in image features before and after fog gas leakage, pixel-level fog gas leakage features including transmittance, coherence and uniformity are designed. Combining the decision-making mechanism of support vector machine and the idea of exponential smoothing, we introduce confidence optimization and smoothing factor constraints to integrate the weights of the algorithm, construct an adaptive foggy gas leakage classification model, and realize the accurate determination of the foggy gas leakage state in the visible light image. The results of the validation of the method in a typical leakage scenario show that the method can better determine the gas leakage state, with a success rate of more than 95%, and the average time consumed for a single frame is about 35 ms, which meets the real-time requirements of detection.