Abstract: In order to realize automated detection of volatile organic compounds (VOCs) based on infrared video, a collaborative modeling method for leakage source location and air plume instance segmentation is proposed. This not only ensures that this model can distinguish gas plume instances, but also ensures that each instance only predicts one leak source. It supports a single branch network to simultaneously achieve the leakage source location and instance segmentation through a single forward inference. Considering the escape characteristics of gas plumes near the leakage source, the location of the leakage source is used as the clustering center of the gas plumes in the embedding space. Based on the spatiotemporal distribution of the leakage gas plume, the covariance variables of the Gaussian distribution probability density function are selected to cluster the pixels in the embedding space and the leakage source location and instance segmentation results for different instances are obtained. The leakage source localization problem is defined as a key point detection problem with a single key point and quantitative evaluation indicators are given. In addition, more precise and accessible annotations are obtained by synthesizing the dataset. Experimental results show that the proposed method can accurately locate the leakage source and segment instances of the gas plume, with higher comprehensive quantitative index than other similar methods, and has good generalization in real videos.