Autonomous mobile robots (AMRs) are increasingly deployed across industrial automation, logistics, healthcare, surveillance and hazardous environments. Their effectiveness depends on the ability to detect and recognise obstacles reliably in unstructured environments. Conventional convolutional neural networks (CNNs) trained from scratch require large annotated datasets and long training times, which limits deployment on resource-constrained mobile platforms. This work studies the modelling of a four-wheel holonomic autonomous mobile robot and applies a deep transfer learning (DTL) algorithm based on a fine-tuned AlexNet CNN for obstacle detection and recognition. The Dynamic Systems Development Model (DSDM) was adopted, and the algorithm was implemented and validated in Simulink/MATLAB Robotics Operating System (ROS). Experimental results across ten test runs yielded a mean recognition accuracy of 98.7% (0.96 s mean detection delay and 97.7% manoeuvre success rate), representing a 1.89% improvement over the baseline CNN benchmark and outperforming several state-of-the-art comparators.