First-person-view (FPV) interceptor drones require stable target tracking under rapid ego-motion, measurement noise, and constrained onboard compute. Image-only methods cannot separate ego-motion from target motion, while classical Kalman filters with fixed covariances lose consistency under rapid viewpoint and scale changes. TAKF is a telemetry-augmented Kalman filtering framework that integrates high-rate attitude telemetry for ego-motion compensation, distance-adaptive noise scaling, and camera-aware covariance transformations. Evaluated on 29 synthetic FPV interception scenarios and real-world flight footage, TAKF achieved higher success rate and mIoU and lower localization error than eight classical and learning-based baselines while maintaining real-time execution on embedded hardware. These results support geometry-consistent filtering for FPV-to-FPV tracking under embedded constraints.