Moving objects introduce observations that conflict with the static-scene model used by LiDAR-inertial-visual odometry (LIVO), affecting both pose estimation and map construction. Semantic detections help identify these observations, but object category alone does not establish motion, and image-plane overlap does not establish camera–LiDAR correspondence. This paper presents RT-LIVO, which links semantic-motion classification, depth-aware measurement selection, and visual weighting within a tightly coupled LIVO framework. Category priors and optical-flow verification assign candidate states, with temporal smoothing and adaptive expansion accounting for intermittent detections and image-plane displacement. The resulting mask controls new visual map-point admission. LiDAR measurements are rejected using image-plane and depth agreement when candidate depth is available, with a two-dimensional fallback otherwise. An empirical reliability score, combining recent valid-feature availability with moving and uncertain candidate ratios, adjusts the photometric covariance in the sequential error-state iterated Kalman filter. On M3DGR, RT-LIVO achieves absolute trajectory error (ATE) and relative pose error (RPE) comparable to the LiDAR-based baselines, while the displayed maps contain fewer dynamic ghost structures. On three self-collected sequences without full-trajectory ground truth, the average endpoint-to-startpoint drift ratio decreases from 4.0257% for FAST-LIVO2 to 0.0793%, indicating improved terminal closure consistency. The average processing time is 25.10 ms per frame on the reported platform. The results support measurement selection and visual weighting for limiting dynamic-scene interference.