Exploiting Spatiotemporal Properties for Efficient Event-Driven Human Pose Estimation