Parking is a pressing issue in high-density areas such as hospitals and universities, particularly in major Asian cities where motorcycles are the predominant mode of transport. These parking facilities are often unstructured and lack demarcation lines, making efficient space management a significant challenge. This paper proposes a deep learning-based method for detecting vacant spots in non-delimited motorcycle parking lots using the U-Net architecture with a ResNet-34 backbone. A custom dataset of 1200 images, captured under diverse real-world conditions, was manually annotated with three semantic labels: “Background,” “Occupied,” and “Vacant.” The model achieves a Mean IoU of 0.7036, an Overall Pixel Accuracy of 86.41%, and a recall of 93.54% for the Vacant class, outperforming Segformer (mIoU: 0.6216, Vacant Recall: 75.62%) in key metrics. Comparative analysis and optimization for edge devices demonstrate its suitability for real-time applications. This work contributes a validated framework and dataset for intelligent parking solutions, advancing smart city initiatives.