Human Detection and Tracking in Frequency Modulated Continuous Wave Radar Point Clouds Using Adaptive Density-Based Spatial Clustering of Applications with Noise

2.1 Proposed method

This study proposed a real-time human tracking system using an IWR6843ISK FMCW radar for outdoor environments. The radar detected human movement and generated three-dimensional point cloud data (x, y, z). Since the tracking is visualized in a top-view perspective, only the (x, y) coordinates are used for planar position estimation.

Before clustering, the point cloud was filtered using Signal-to-Noise Ratio (SNR) and Doppler velocity thresholds. This filtering step removes low-quality detections and static reflections, thereby reducing noise and ghost targets. The filtered 2D point cloud is clustered using DBSCAN to separate true targets from residual noise. Each cluster’s centroid is then tracked using a KF, which performs predictive state estimation and measurement updates. This combination of filtering, density-based clustering, and predictive tracking enables accurate, stable, and robust performance in complex outdoor scenarios.

Beyond the overall system workflow shown in Figure 1, the adaptive DBSCAN module follows an internal processing sequence designed to adjust the clustering parameters based on range-dependent point cloud characteristics. This procedure is summarized in Figure 2, which outlines how the radar points are filtered, analyzed, and assigned to distance zones before the appropriate DBSCAN parameters are applied. The same workflow is also used during real-time operation, where each incoming frame is processed through this sequence to ensure that the adaptive ϵ and Minimum Points (MinPts) configuration remains effective under dynamic conditions and varying target ranges.

Figure 1. Proposed method

Figure 2. Adaptive Density-Based Spatial Clustering of Applications with Noise (DBSCAN) processing workflow

The experiments were conducted to assess the performance of the proposed FMCW radar-based tracking system, with the primary focus placed on evaluating the adaptive DBSCAN clustering mechanism. The analysis examines how the adaptive zoning strategy improves cluster formation across varying distances and contributes to stable trajectory estimation when combined with the KF. In addition to algorithmic evaluation, the experiments also verify the radar’s tracking capability by measuring its effective detection range in real-time and comparing the estimated trajectory against a predefined reference path using Root Mean Square Error (RMSE).

The performance of the DBSCAN algorithm is highly sensitive to the choice of its parameters, namely the neighborhood radius ϵ and the minimum number of points (MinPts). This sensitivity becomes a major challenge in radar point cloud data, where the spatial density of detections is non-uniform. Near targets tend to produce denser point returns due to stronger reflected power. In contrast, distant targets yield sparser point clouds as a result of propagation loss and signal attenuation [3]. As a consequence, a single set of DBSCAN parameters cannot consistently yield optimal clustering results across different ranges.

To address the density variation across distance, the behaviour of the collected point cloud data was first examined. Figure 9 shows the point density as a function of range, revealing a high-density region at short distances, a transitional region in the mid-range, and a sparse region beyond approximately 30 m.

Figure 9. Point density vs. range

This pattern is further confirmed in Figure 10, where the per-frame density plotted against the median range forms three naturally separated bands. These observations indicate that the measurement space does not vary uniformly but instead exhibits three distinct density regimes.

Figure 10. Density vs. range per frame

Based on this empirical evidence, a three-zone division was adopted, and K-Means clustering applied to the range density pairs produced the numerical boundaries summarized in Table 1. This zoning allows the DBSCAN parameters to adapt dynamically to the spatial density of the detections, improving clustering reliability across different ranges.

Table 1. Distance zone

Table 2. Parameter search range per zone

The parameter search ranges for each zone are listed in Table 2, where ε and MinPts were systematically varied using a grid search procedure. Since the dataset contains a single moving target, the quality of each parameter configuration was evaluated using a Cluster Count Accuracy (CCA), which measures how consistently DBSCAN produces exactly one cluster per frame, expressed as Eq. (7).

CCA (%)$=\left( \frac{\mathop{\sum }_{i=1}^{N}1\left( {{C}_{i}}=1 \right)}{N} \right)\times $100%                            (7)

where,

N: the total number of frames;

C_i: the number of clusters produced in frame i;

$1\left( {{C}_{i}}=1 \right)$: an indicator function that returns 1 if frame $i$ contains exactly one cluster, and 0 otherwise.

The parameter set yielding the highest proportion of correct single-cluster detections was selected as the optimal configuration for each zone.

Parameter tuning was performed using a grid search procedure, and each configuration was evaluated based on how consistently DBSCAN produced a single cluster per frame. The best performing settings for each zone are summarized in Table 3. The near zone yielded the highest stability due to its dense point returns, whereas the medium and far zones required larger ε values to accommodate the reduced point density at longer distances. These results confirm that the zone-based adaptation enables DBSCAN to maintain reliable clustering behaviour across varying ranges.

Table 3. Optimal Adaptive Density-Based Spatial Clustering of Applications with Noise parameters per zone

This adaptive zoning strategy substantially improves the robustness of the clustering process by ensuring that the DBSCAN parameters are matched to the underlying density characteristics of each range segment. By doing so, it mitigates over-segmentation in the dense near field region and reduces the risk of missed detections in the sparser mid and far field regions. Consequently, the algorithm is able to generate stable and consistent single-cluster outputs across all distances. The optimized clustering results from each zone are then used as inputs to the KF for trajectory estimation, as discussed in the following subsection.

Before proceeding to trajectory estimation with the KF, the clustering performance of the adaptive DBSCAN algorithm was evaluated using the CCA, which measures how consistently the algorithm produces a single valid cluster per frame. Two experimental scenarios were considered: horizontal movement at distances ranging from 5 to 50 m, and radial movement in which the subject approached and receded along the Y-axis.

Table 4 shows that clustering performance is highest at close range, with CCA values of 95.74% at 5 m and 93.40% at 10 m. Accuracy gradually decreases as the subject moves farther away, reaching the mid 80% range at distances between 15 and 30 m. A more noticeable decline appears beyond 35 m, where accuracy drops to 82.51% at 35 m and 80.67% at 40 m. The performance then degrades sharply at the farthest distances, falling to 50.77% at 45 m and 33.05% at 50 m. This sharp decline is caused not only by increased noise but also by a significant loss of detections as the radar approaches its effective range limit. With fewer valid points available per frame, DBSCAN becomes increasingly difficult to maintain a consistent single-cluster output.

Table 4. Evaluation at horizontal distances

For the radial movement scenario, the adaptive DBSCAN also produced consistent clustering results, as shown in Table 5. When the subject moved toward the radar, the algorithm achieved an accuracy of 87.50%, while receding motion produced a slightly higher accuracy of 91.81%. These results indicate that the zone-based approach can effectively handle changes in point cloud distribution caused by different motion directions, allowing the clustering to remain reliable in both horizontal and radial movements.

Table 5. Cluster Count Accuracy evaluation in radial movements

Overall, the obtained results validate that the adaptive DBSCAN provides stable and accurate clustering across varying distances and movement patterns. However, despite the reliable formation of clusters, the centroid positions still exhibited minor frame-to-frame fluctuations due to noise and variations in the radar point cloud. To mitigate this effect and enhance temporal consistency, the KF was applied to refine the centroid trajectories, aiming to produce smoother and more stable motion estimation while preserving target identity.

In this stage, centroids from valid clusters were treated as measurement inputs for the KF-based tracking module. The KF performed recursive prediction and update steps for each detected object, enabling continuous tracking while maintaining consistent identity through frame-to-frame association. The comparison presented in Tables 6 and 7 highlights this improvement. The raw centroids obtained from DBSCAN showed slight positional instability across consecutive frames, whereas the KF output produced smoother and more continuous trajectories with consistent Track IDs.

As shown in Tables 6 and 7, the raw centroids derived from DBSCAN exhibit minor frame-to-frame fluctuations caused by variations in the point cloud distribution. After KF filtering, the estimated trajectory becomes noticeably smoother and more stable, with the same track ID maintained throughout the sequence. This confirms that the KF effectively performs temporal smoothing and noise suppression while preserving object identity. Figure 11 further visualizes these effects, illustrating how the KF output provides a more coherent and continuous trajectory compared to the unfiltered clustering results.

Table 6. Centroid positions before tracking

Table 7. Centroid positions after Kalman filter tracking

(a)

(b)

Figure 11. Trajectories before and after Kalman filter (KF) tracking: (a) Without KF; (b) With KF

To further quantify the tracking accuracy, the KF estimated trajectories were compared against predefined reference paths under controlled angular movements. To assess the overall tracking accuracy, the trajectories were evaluated against linear reference paths defined at nine azimuth angles ranging from -60° to 60°, in 15° increments. Each reference path was modeled as a straight line $y=mx$, where $m=\text{tan}\left( \theta  \right)$ and the radar was positioned at the origin (0,0). Figure 12 illustrates an example of the test setup at an azimuth angle of -15°, where the detected centroid trajectory is evaluated relative to the corresponding reference line.

Figure 12. Example of trajectory evaluation at -15° reference angle

For each frame, the perpendicular distance from the centroid to its corresponding reference line was calculated. The Mean Error and RMSE were then used as quantitative performance metrics, summarizing tracking accuracy across all azimuth directions. The results are presented in Table 8.

Table 8. Tracking accuracy across reference angles