Continuous gait monitoring is essential for preventive healthcare and fall-risk assessment. Existing wearable devices are often constrained by high cost, material rigidity, or limited single-modality discrimination, which hinders their ability to distinguish biomechanically similar activities such as slow walking and stair climbing. This study presents a technical proof-of-concept for an embedded smart footwear system that integrates low-cost (<$2) paper-based graphene pressure sensors (PBGS) with dual inertial measurement units (IMUs). The PBGS are produced via a scalable aqueous dip-coating process and achieve high linearity (R² > 0.92) across a unified 0–5 kg (0–122.6 kPa) calibration envelope; extended characterisation to 10 kg and a per-subject dynamic analysis confirm that fast-walking heel peaks reach 1.35–1.55 × the 5 kg reference within the unsaturated range of the sensor. The system synchronises four-channel kinetic pressure and twelve-axis kinematic data at a nominal 100 Hz. Recognition of four instructed activities (standing, slow walking, fast walking, stair climbing) was evaluated using a strict Leave-One-Subject-Out (LOSO) protocol across 15 subjects. The XGBoost pressure-only model achieved a mean accuracy of 78.64% and a macro F1-score of 0.6765, whereas probability-level fusion of the pressure and IMU branches increased performance to 88.83% accuracy and a macro F1-score of 0.8397, corresponding to improvements of 10.19 %age points and 0.1632, respectively. XGBoost outperformed SVM, Random Forest and LightGBM baselines, as well as 1D-CNN and LSTM models trained on raw windows. These results support the complementarity of plantar-pressure and inertial sensing for subject-independent activity recognition, pending validation in larger and clinical cohorts.