Ensuring the safety of motorcycle passengers, particularly pillion riders, remains a persistent challenge within rapidly growing smart cities, given the high vulnerability of these users in traffic accidents and the limited deployment of real-time protection technologies. This study presents a comprehensive end-to-end framework for real-time fall detection and spatial risk prediction, integrating helmet-embedded inertial measurement units (IMUs), an edge-based hybrid CNN–LSTM deep learning model optimized with Artificial Bee Colony (ABC-Scout) algorithms, and geospatial risk mapping via ordinary kriging. The system was validated using a heterogeneous dataset of 1,500 annotated urban riding and controlled fall events, reflecting diverse surface and environmental conditions in Ciudad Juárez. Results demonstrate robust classification performance (accuracy: 97.8%, F1-score: 97.3%, recall: 98.2%, precision: 96.5%), operationalized with an average inference latency below 50 m on embedded platforms. Spatial modeling revealed high-risk urban corridors, strongly correlating with historical accident hotspots. The proposed solution enables continuous, scalable and privacy-aware monitoring, supporting both instant passenger protection and strategic data-driven interventions in smart city environments.