Event-Based Bidirectional People Counting Using a Spatiotemporal Neural Network and a Single 77 GHz MIMO FMCW Radar.

Abstract

Accurate bidirectional people counting at room entrances is essential for reliable occupancy estimation, yet radar-based methods remain challenged when reflections from multiple individuals temporarily merge or when people move simultaneously in opposite directions. This study presents a hybrid event-based method for bidirectional people counting using a single 77 GHz multiple-input multiple-output frequency-modulated continuous-wave radar. Unlike conventional tracking-based approaches, the proposed method treats a complete doorway passage as one event and does not require continuous individual trajectories or frame-to-frame data association. Each event is represented as a sequence of three-dimensional point clouds with radial velocity information. A PointNet encoder extracts frame-level spatial representations, a bidirectional gated recurrent unit models their temporal evolution, and an attention mechanism emphasizes the most informative event phases. Separate output heads estimate the numbers of entering and exiting individuals, while an auxiliary head identifies the event type. Predictions are accepted using explicit confidence and geometric-consistency criteria. The primary dataset comprised 550 labeled events and an additional 100 events recorded in five continuous sequences. In the held-out test session, entry- and exit-count accuracies were 95.5% and 94.5%, respectively, while joint entry-exit count accuracy reached 92.7%. The mean absolute error in occupancy change was 0.06 person per event, and the mean absolute occupancy error across the continuous sequences was 0.10 person. The proposed method exceeded geometric rules and density-based clustering combined with tracking in joint count accuracy by 37.2 and 24.5 percentage points, respectively. On an independent laboratory dataset comprising 100 events, the model achieved a joint count accuracy of 91.0% without retraining, fine-tuning, or parameter adaptation, providing preliminary evidence of cross-environment generalization.