Infrared (IR) image processing is used for military applications where performances focus on the earliness of detection, on the tracking of a target and on the reduction of the false alarm rate. In the IR context, classical detect-before-track(DBT) algorithms may be unreliable because they first rely on thresholding the image to obtain plots or detections which are next tracked which implies a loss of information. Detecting low signal to noise ratio (SNR) targets (such as faraway targets) requires decreasing the detection threshold. Consequently, the false alarm rate and missed detections may increase. Moreover, the data association problem which consists in associating plots with targets is computational intensive. Track-before-detect (TBD) algorithms do not have these drawbacks because they use raw measurements rather than detections. In this paper, we adapt classical TBD algorithms used in radar for the IR context. We next compute both a DBT algorithm and a TBD algorithm and we compare these methods in the context of very low SNR targets in IR image sequences