The assessment of cracks in civil infrastructures commonly relies on visual inspections carried out at night, resulting in limited inspection time and an increased risk of crack oversight. The Digital Image Correlation (DIC) technique, employed in structural monitoring, requires stationary cameras for image collection, which proves challenging for long-term monitoring. This paper describes the Crack Monitoring from Motion (CMfM) methodology for automatically detecting and measuring cracks using non-fixed cameras, combining Convolutional Neural Networks and photogrammetry. Through evaluation using images obtained from laboratory tests on concrete beams and subsequent comparison with DIC and a pointwise sensor, CMfM demonstrates accurate crack width computation within a few hundredths of a millimetre when compared to the sensor. This method exhibits potential for effectively monitoring temporal crack evolution using non-fixed cameras.