Infrared imagery has been used in many areas such as military, surveillance, medical imaging and numerous industrial branches. The recent increase in the use of infrared (IR) imaging techniques in various fields draws the attention of more studies towards this area. One of the common problems of the thermal IR detector units is the existence of bad pixels. Bad pixels may arise from various reasons such as manufacturing processes or operating conditions. This phenomenon is commonly dealt with well-known calibration methods. However, they are generally applied at factory level or they interrupt the operational use due to the need for utilizing a uniform reference scene. For those reasons recent methods employ scene based approaches without requiring a special equipment. Those types of methods commonly make some assumptions based on statistical characteristics of bad pixels. They mainly assume that bad pixels deviate at a certain level from their neighboring pixels. They rely on sufficient variation of scene content in time and the fact that possible false detections can be canceled out due to scene variation. Nevertheless, this assumption does not always hold, especially when the camera is stationary. In such cases, some distinctive parts of the underlying scene may be falsely regarded as bad pixels. To that end, we develop a method that is able to isolate the scene content from bad pixels in order to eliminate erroneous detections of scene parts. The proposed method benefits from the motion of the camera which provides responses of different pixels for the same scene region. From this information, we expect similar responses for the registered pixels, if they are not defective. On the other hand, if the pixel responses are exceedingly different, then we can deduce that the corresponding pixel may be defective. For this purpose, we first register adjacent frames using an efficient 1D projection based matching method. To ensure a more robust registration, we use edge maps rather than the intensity image. After the registration of two frames, we construct an error map for the overlapping regions of the two frames. We declare our candidate defective pixels by assessing the deviation levels of error values. Candidate pixels are accumulated across non-stationary frames to obtain temporally consistent detections. Since our inter-frame registration step provides motion information, we avoid accumulation when camera is stationary. We also prevent erroneous registrations by checking for the sufficient scene detail. The performance evaluations are carried out on an extensive dataset consisting of real thermal camera images. The dataset contains a wide variety of scene content and various scenarios featuring stationary camera conditions that causes failures in traditional statistical variation based approaches. The results of our experiments are assessed in terms of true and false bad pixel detections as compared to ground-truth bad pixel labellings. The results show that the proposed inter-frame registration based bad pixel detection method achieves successful results without any assumption about scene content and any additional reference surface.