In order to solve the on-line detection of the body surface temperature for sow based on thermal infrared video, the image segmentation method for the fast and efficient target detection was proposed. The thermal infrared image of the sow has the features of low pixel, low contrast and edge blur. In piggery environment conditions, the sow body temperature and background radiance were the main factors to affect thermal infrared image brightness and handling results. Because of the strong correlation between the intensity of the background radiation and light intensity, in order to study the effect of background radiation on the thermal infrared image segmentation, the thermal infrared images of different illumination intensities were collected. Firstly, the point operation was used to enhance the contrast enhancement; and then, instead of a constant value for ω , a weight function that varies dynamically with the global and local contrast of the image was chosen, so as to dynamically balance the global energy and the local energy; finally, an improved LGIF model was established with the global fitting energy and the local energy. 300 thermal infrared images were collected by using infrared thermal imaging system, and the image segmentation experiments were performed. These pictures were taken in different positions, light conditions, and sow varieties. Classification tests were carried out under three conditions of low light intensity (100~600 lx), middle illumination (600~1 000 lx) and high illumination (1 500~2 500 lx). In order to analyze the accuracy and real-time performance of the algorithm, the average running time and the correct segmentation rate of different segmentation algorithms were calculated respectively. The cause of the poor effect of the partial sample was analyzed, and the direction of improvement was put forward. Experimental results show that the improved method can extract the sow more efficiently, and the average single image segmentation time was 49.67 s, the correct segmentation rate reached more than 98% which demonstrated the accuracy and superiority of the proposed model.