This article introduces a new method to improve the dependability of millimeter-wave (mmWave) and terahertz (THz) network services in dynamic outdoor environments. In these settings, line-of-sight (LoS) connections are easily interrupted by moving obstacles like humans and vehicles. The proposed approach, coined as Radar-aided Dynamic blockage Recognition (RaDaR), leverages radar measurements and federated learning (FL) to train a dual-output neural network (NN) model capable of simultaneously predicting blockage status and time. This enables determining the optimal point for proactive handover (PHO) or beam switching, thereby reducing the latency introduced by 5G new radio procedures and ensuring high quality of experience (QoE). The framework employs radar sensors to monitor and track objects movement, generating range-angle and range-velocity maps that are useful for scene analysis and predictions. Moreover, FL provides additional benefits such as privacy protection, scalability, and knowledge sharing. The framework is assessed using an extensive real-world dataset comprising mmWave channel information and radar data. The evaluation results show that RaDaR substantially enhances network reliability, achieving an average success rate of 94% for PHO compared to existing reactive HO procedures that lack proactive blockage prediction. Additionally, RaDaR maintains a superior QoE by ensuring sustained high throughput levels and minimising PHO latency.