In elite track cycling, performance gains increasingly depend on event-level analysis of training and competition data. However, current practices rely heavily on manual tasks, limiting scalability and real-time feedback. This paper presents an integrated system for team pursuit analysis that unifies multimodal sensor data (timing loops, athlete-worn sensors) and automatic video annotations from pan-tilt-zoom footage within a centralized Wireless Cycling Network (WCN). Sensor and timing data are leveraged to generate rule-based annotations for key events, such as lead switches, which serve as ground truth labels for training vision models. A RF-DETR model is fine-tuned on a diverse track cycling dataset to detect track cyclists and is integrated into a semantic tracking pipeline specific to the team pursuit formation. Instead of continuous identity tracking, the method exploits domain-specific priors on rider order and train geometry to detect lead transitions. Evaluations across training and competition footage demonstrate reliable cyclist detection (F1-score up to 0.96 at IoU 0.50) and lead switch detection (95%). By combining both sensor and video based data, the system can self-annotate new data and provide a unified analysis interface for both training and races while simultaneously lowering the manual input for coaches.