Multimodal approaches in hazard detection and risk assessment: current methods and future trends

Hazard detection and risk assessment are important across transportation, healthcare, mining, industry, and environmental monitoring. Advances in sensor technology and deep learning have increased the use of multimodal approaches that combine complementary sensing and data sources. This structured narrative review examines 89 publications and classifies hazard detection methods as visual-based, environmental sensor-based, wearable device-based, or multimodal fusion methods. Risk assessment methods are organized as deterministic and rule-based, probabilistic, statistical and data-driven, or methods under non-statistical uncertainty. Among the 39 original hazard detection studies, 27 (69.2%) used single-modal or non-fusion approaches, whereas 12 (30.8%) used explicit multimodal fusion. Representative within-study comparisons showed that camera–LiDAR fusion increased detection accuracy from 58.3% to 74.7%, while radar- camera fusion reduced the average false-negative rate from 6.46% for the camera and 6.59% for the radar to 0.61% for the fused tracker. These findings indicate that multimodal fusion can improve detection robustness within the same evaluation setting, although the magnitude of improvement depends on the task, dataset, sensors, and fusion architecture. The review also finds that risk assessment method selection depends mainly on data availability, uncertainty type, explainability, computational resources, and response-time requirements. Sensor alignment, data quality, real-time processing, maintenance cost, and limited common evaluation procedures remain major challenges.