The use of personal transportation devices such as e-bikes and e-scooters (micromobility) necessitates the development of improved safety support systems using highly-accurate, real-time lane recognition. However, the constrained operating environment, both computationally and physically, on such devices restricts the applicability of existing sensor-based solutions. One option is to leverage vision-based systems and AI models. However, these are typically built using high-spec processors and high-memory platforms and the models need to be adapted to low-spec platforms such as microcontrollers. A significant barrier to the development and evaluation of these potential solutions is the lack of lane recognition datasets that focus on the first-person (rider) perspective. We contribute a lane recognition dataset of micromobility first-person perspective images from e-mobility rides. This dataset is utilized to assess the impact of channel and spatial attention on compact CNN models, driven by the aim to maximize utilization through the addition of cost-effective operations like these attention mechanisms, which introduce only a modest increase in the number of parameters. We find that adding channel and spatial attention can improve the performance of the standard compact CNN classification models and specifically that adding the spatial branch improves the performance of the model with channel attention. The MobileNetV3 model with the fewest parameters among those with channel plus spatial attention maintained high overall performance. Our code and dataset are publicly accessible at: https://github.com/Luna-Scooters/Compact-Attention-based-CNNs-on-MLRD.