FARSEEING Real-World Falls Dataset

• FARSEEING is a real-world, clinically validated falls repository capturing genuine fall events in older adults through wearable tri-axial inertial sensors.
• The dataset aggregates heterogeneous sensor data (accelerometer, gyroscope, magnetometer) from diverse placements and sampling rates to support realistic fall detection studies.
• Its evaluation framework emphasizes continuous monitoring, addressing class imbalance and event-level detection to mirror real-world deployment conditions.

Searching arXiv for the papers on arXiv and related FARSEEING material. Searching arXiv for "(Schneider et al., 26 May 2025) FARSEEING OmniFall" Searching arXiv for "(Aderinola et al., 15 Sep 2025) FARSEEING real-world falls dataset" The FARSEEING Real-World Falls Dataset is a real-world falls repository centered on clinically verified falls recorded with wearable tri-axial inertial sensors in older adults and other clinically relevant fallers. In recent fall-detection research, it is treated as a rare source of non-simulated fall data collected in free-living or clinical contexts rather than as a staged benchmark, and it is therefore used to study deployment-relevant behavior under continuous monitoring, class imbalance, and heterogeneous sensing conditions (Aderinola et al., 15 Sep 2025). Within the broader landscape of fall detection datasets, FARSEEING occupies a distinct position: it is outside the OmniFall benchmark, which unifies eight primarily staged video datasets and a separate real-world subset curated from OOPS!, and it is characterized instead by ecological and clinical validity in home and care settings (Schneider et al., 26 May 2025).

1. Dataset identity and scope

FARSEEING is described as the FARSEEING real-world falls repository. Its defining property is that it contains real-world falls, not simulated, and that the falls are clinically verified and annotated (Aderinola et al., 15 Sep 2025). The population is explicitly older-adult and clinically oriented. In the accelerometer-based streaming study, the reported cohort comprises 92 fallers, with mean age $76.1 \pm 12.6$ years, and the repository overall is summarized as containing 208 clinically verified falls with inertial sensors (Aderinola et al., 15 Sep 2025).

This clinical and ecological orientation distinguishes FARSEEING from staged fall corpora. OmniFall’s eight component datasets—CMDFall, UP-Fall, Le2i, GMDCSA-24, EDF, OCCU, MCFD, and CAUCA Fall—are primarily staged, whereas FARSEEING is described as a real-world, clinical/home monitoring falls corpus containing genuine falls of older adults recorded in care facilities and homes (Schneider et al., 26 May 2025). FARSEEING is not one of the eight staged datasets unified in OmniFall, and it is not the source of OOPS-Fall; OmniFall’s real-world subset is curated from the OOPS! accident-video dataset instead (Schneider et al., 26 May 2025).

A recurring theme in recent work is that FARSEEING is not simply another dataset instance but a test of whether fall detection methods remain valid when controlled assumptions are removed. The emphasis on clinically verified, real-world falls is used to argue that models trained and evaluated only on simulated or pre-segmented falls do not directly establish deployment reliability (Aderinola et al., 15 Sep 2025).

2. Sensor modalities, placement, and recording structure

FARSEEING is heterogeneous because it aggregates data across multiple studies. The repository uses wearable tri-axial inertial sensors, with different combinations of accelerometer, gyroscope, and magnetometer, and with different placements and sampling rates (Aderinola et al., 15 Sep 2025). The placements explicitly mentioned are L5 (lumbar, lower back) and thigh. The sampling rates explicitly mentioned are 20 Hz and 100 Hz (Aderinola et al., 15 Sep 2025).

A recent streaming framework restricts analysis to a more homogeneous subset: 145 falls from 41 participants from the FARSEEING dataset with sensors placed at the L5 position and with sampling rates of 100 Hz, to maintain consistency and ensure high data quality (Aderinola et al., 15 Sep 2025). That restriction is methodological rather than intrinsic to the full repository. It indicates that FARSEEING’s heterogeneity is sufficiently large that many studies may need to define controlled subcohorts before model development.

The temporal structure of the recordings used in that study is also explicit. Each fall recording contains 20 minutes of accelerometer data, with the fall impact event at the 10th minute, yielding a symmetric structure of 10 minutes pre-fall and 10 minutes post-fall (Aderinola et al., 15 Sep 2025). The same work reports using over 60 hours of IMU data in total, which combines fall-centered recordings and additional non-fall continuous segments used to generate negative samples (Aderinola et al., 15 Sep 2025).

The following summary captures the dataset properties explicitly described in the recent literature.

Aspect	FARSEEING characteristic	Source
Fall type	Real-world, clinically verified, annotated	(Aderinola et al., 15 Sep 2025)
Repository scale	208 clinically verified falls with inertial sensors	(Aderinola et al., 15 Sep 2025)
Cohort used in streaming study	145 falls from 41 participants; 92 fallers, mean age $76.1 \pm 12.6$ years	(Aderinola et al., 15 Sep 2025)
Sensor placement	L5 and thigh	(Aderinola et al., 15 Sep 2025)
Sampling rates	20 Hz and 100 Hz	(Aderinola et al., 15 Sep 2025)
Recording structure in streaming study	20 minutes per signal; impact at the 10th minute	(Aderinola et al., 15 Sep 2025)

In comparison with video-based fall datasets, FARSEEING is likely single- or few-view and noisier, with less controlled lighting, while remaining clinically more representative than staged datasets (Schneider et al., 26 May 2025). That comparison is conceptual rather than a statement about a single uniform acquisition protocol, because the repository itself is sensor-heterogeneous.

3. Annotation model and signal representation

A key feature of recent FARSEEING-based work is the use of a multiphase fall model derived from Becker et al. (2012). The fall event is decomposed into three phases: falling phase $[t_0, t_1)$, impact phase $[t_1, t_2)$, and post-fall phase $[t_2, t_3)$, with total duration

$w = t_3 - t_0.$

In the training segmentation used for accelerometer-based streaming detection, the impact annotation supplied by FARSEEING is taken as $t_1$, and the surrounding boundaries are defined as $t_0 = t_1 - 1$ second and $t_2 = t_1 + 1$ second (Aderinola et al., 15 Sep 2025). Each training fall segment is therefore

$[t_0, t_3) = [t_1 - 1, t_1 - 1 + w),$

which contains 1 second of pre-impact falling, 1 second of impact, and $76.1 \pm 12.6$0 seconds of post-fall behavior (Aderinola et al., 15 Sep 2025). The paper explicitly treats the post-fall phase as important for separating true falls from near-falls.

The sensor signal is reduced from tri-axial accelerometry to a univariate magnitude,

$76.1 \pm 12.6$1

where $76.1 \pm 12.6$2, $76.1 \pm 12.6$3, and $76.1 \pm 12.6$4 denote the anterior-posterior, medial-lateral, and vertical axes, respectively (Aderinola et al., 15 Sep 2025). This magnitude representation is used because it reduces sensitivity to sensor orientation and allows each segmented sample to be represented directly as a fixed-length vector. The preprocessing steps described are aggregation to magnitude, segmentation, and standardization; the study explicitly states that it performs no feature extraction, and that each segmented sample is represented as a vector (Aderinola et al., 15 Sep 2025).

For a window size $76.1 \pm 12.6$5 seconds at 100 Hz, each segment contains

$76.1 \pm 12.6$6

samples, and the training set is represented as

$76.1 \pm 12.6$7

with $76.1 \pm 12.6$8 for fall windows and $76.1 \pm 12.6$9 for ADL windows (Aderinola et al., 15 Sep 2025). This formulation makes FARSEEING compatible with standard probabilistic classifiers over fixed-length univariate time series.

4. Negative sampling, class imbalance, and streaming evaluation

FARSEEING includes substantial non-fall data, operationalized as activities of daily living (ADLs) in the streaming study. Negative samples are generated from continuous segments not labeled as falls using fixed-size overlapping sliding windows of length $[t_0, t_1)$0 and step size 1 second (Aderinola et al., 15 Sep 2025). To avoid populating the negative class with only inactivity, the study applies a signal-amplitude filter and retains only windows where the maximum acceleration magnitude exceeds 1.4 g, citing earlier FARSEEING work indicating that real fall impacts typically produce peaks above that value (Aderinola et al., 15 Sep 2025).

The resulting training distribution is explicitly imbalanced. Example folds contain approximately 602–1021 ADL windows and 86–113 fall windows, while the final training set contains 1089 ADL windows and 124 fall windows (Aderinola et al., 15 Sep 2025). The paper characterizes this as clear imbalance, with ADLs outnumbering falls by roughly an order of magnitude.

Evaluation is designed as real-time, continuous monitoring without prior knowledge of fall events. Test signals are left unsegmented; each is a continuous 20-minute recording with one annotated fall (Aderinola et al., 15 Sep 2025). A sliding window of length $[t_0, t_1)$1 and step 1 second is run over each signal, and the classifier outputs a fall probability $[t_0, t_1)$2 for each window $[t_0, t_1)$3. The study then applies an additional rule: for windows where the maximum acceleration in $[t_0, t_1)$4 is less than 1.4 g, the probability is set to zero (Aderinola et al., 15 Sep 2025). Window-level probabilities are aggregated into a sample-level confidence map via

$[t_0, t_1)$5

This produces a continuous probability trace aligned with the raw signal (Aderinola et al., 15 Sep 2025).

Event-level detection is defined with an asymmetric tolerance window around the impact index $[t_0, t_1)$6. The ground-truth fall event is

$[t_0, t_1)$7

and the tolerance window is

$[t_0, t_1)$8

with $[t_0, t_1)$9 seconds (Aderinola et al., 15 Sep 2025). For a detection window $[t_1, t_2)$0, the criterion is based on

$[t_1, t_2)$1

A detection is a true positive if $[t_1, t_2)$2, a false positive if $[t_1, t_2)$3, and a false negative if no detection overlaps $[t_1, t_2)$4 (Aderinola et al., 15 Sep 2025). This definition is explicitly intended to be more realistic than per-sample labeling because it evaluates whether each fall triggers an acceptable alarm.

A common misconception in this area is that very high performance on pre-cut fall windows directly implies usefulness in continuous monitoring. The FARSEEING-based streaming protocol is constructed precisely to avoid that assumption: fall times are unknown at test time, windows are generated online, and evaluation is event-based rather than based on pre-segmented exemplars (Aderinola et al., 15 Sep 2025).

5. Benchmark use, classifiers, and reported results

Recent work on FARSEEING benchmarks five classifiers on the same raw-magnitude representation: ExtraTrees, MiniRocket, QUANT, Catch22, and ResNet (Aderinola et al., 15 Sep 2025). All models use default hyperparameters in their respective libraries, and all are trained and evaluated under participant-wise splits, so that data from a given participant appear in only one set (Aderinola et al., 15 Sep 2025). A 20% hold-out test cohort is reserved as an untouched test set; the remaining 80% is used for model and window-size selection via cross-validation (Aderinola et al., 15 Sep 2025).

The window sizes evaluated are

$[t_1, t_2)$5

For each $[t_1, t_2)$6, the study reports Balanced Accuracy, Precision, Recall, F1, and Detection Delay using an untuned threshold $[t_1, t_2)$7, and then performs cost-sensitive threshold tuning (Aderinola et al., 15 Sep 2025).

Across cross-validation, QUANT is reported as the best overall model, with $[t_1, t_2)$8 seconds and the following mean performance: BA = 0.94 ($[t_1, t_2)$9), Precision = 0.79 ($[t_2, t_3)$0), Recall = 0.87 ($[t_2, t_3)$1), F1 = 0.82 ($[t_2, t_3)$2), and Delay = 0.60 s ($[t_2, t_3)$3) (Aderinola et al., 15 Sep 2025). The same section reports ExtraTrees with BA = 0.90, Precision = 0.77, Recall = 0.79, F1 = 0.78, and delay approximately 2.43 s; MiniRocket with BA = 0.93, Precision = 0.70, Recall = 0.86, F1 = 0.77, and Delay = -6.90 s; Catch22 with BA = 0.90, Precision = 0.55, Recall = 0.80, F1 = 0.64; and ResNet with BA = 0.86, Precision = 0.64, Recall = 0.73, F1 = 0.66 (Aderinola et al., 15 Sep 2025).

On the hold-out test set, the highlighted result is the QUANT model with $[t_2, t_3)$4 seconds. Without cost-sensitive tuning, at $[t_2, t_3)$5, it achieves BA = 0.95, Precision = 0.84, Recall = 0.90, and F1 = 0.87 (Aderinola et al., 15 Sep 2025). With cost-sensitive tuning, using $[t_2, t_3)$6, the reported performance becomes BA = 1.00, Precision = 0.84, Recall = 1.00, and F1 = 0.91, eliminating false negatives while adding one additional false alarm (Aderinola et al., 15 Sep 2025). The same work reports average inference time below 5 ms per sample, interpreting this as compatible with real-time wearable deployment (Aderinola et al., 15 Sep 2025).

The cost-sensitive operating point is defined by the gain matrix

$[t_2, t_3)$7

and the gain function

$[t_2, t_3)$8

with thresholds scanned over 100 evenly spaced values in $[t_2, t_3)$9 (Aderinola et al., 15 Sep 2025). This formulation encodes the explicit assumption that the cost of a missed fall is at least twice that of a false alarm.

The comparison with earlier FARSEEING studies is methodologically important. The streaming study contrasts its results with prior non-streaming or pre-segmented evaluations, including Palmerini et al. (2020), Ramanathan & McDermott, and Aderinola et al. (2024), and argues that high non-streaming metrics may rely on known fall boundaries or pre-segmented windows (Aderinola et al., 15 Sep 2025). This suggests that FARSEEING is not only a dataset but also a testbed for evaluation design.

6. Relation to OmniFall and interoperability with video benchmarks

Although FARSEEING is not part of OmniFall, the OmniFall paper provides a useful external frame for understanding its role. OmniFall unifies eight public fall detection datasets under a consistent ten-class taxonomy with dense frame-level segmentation, but those eight datasets are primarily staged and video-based; FARSEEING is absent from the benchmark (Schneider et al., 26 May 2025). OmniFall’s real-world component, OOPS-Fall, is curated from accident videos on the internet rather than from clinical or home-care monitoring (Schneider et al., 26 May 2025).

The contrast is domain-specific. OOPS-Fall represents a broad “internet” fall domain with arbitrary viewpoints, dynamic framing, sports and everyday accidents, and often younger subjects, whereas FARSEEING is characterized as a clinical/home-care domain involving older adults, indoor environments, care facilities, chronic conditions, and assistive devices (Schneider et al., 26 May 2025). OOPS-Fall is therefore “in the wild” in a different sense from FARSEEING. The literature explicitly notes that models generalizing to OOPS-Fall may still struggle with FARSEEING because subject demographics, camera setups, and motion profiles are qualitatively different (Schneider et al., 26 May 2025).

OmniFall also proposes a ten-class label set

$w = t_3 - t_0.$0

including transient actions such as $w = t_3 - t_0.$1, $w = t_3 - t_0.$2, $w = t_3 - t_0.$3, $w = t_3 - t_0.$4; static states such as $w = t_3 - t_0.$5, $w = t_3 - t_0.$6, $w = t_3 - t_0.$7, $w = t_3 - t_0.$8; and the additional classes $w = t_3 - t_0.$9 and $t_1$0 (Schneider et al., 26 May 2025). The paper explicitly describes how FARSEEING could be mapped into this taxonomy: the collapsing phase would be labeled as $t_1$1, time on the floor after a fall as $t_1$2, and pre-fall activities such as walking, sitting, or lying as $t_1$3, $t_1$4, or $t_1$5, respectively (Schneider et al., 26 May 2025). If fine-grained transitions are unavailable, they can be inferred from pose changes or treated as $t_1$6 (Schneider et al., 26 May 2025).

That proposed alignment remains hypothetical rather than native to the repository. FARSEEING’s raw annotations are described as typically marking fall events and possibly fallen intervals more coarsely, whereas OmniFall requires dense frame-level segmentation (Schneider et al., 26 May 2025). A plausible implication is that making FARSEEING directly comparable to OmniFall would require substantial relabeling or semi-automatic segmentation.

7. Research significance, limitations, and methodological implications

FARSEEING is significant because it anchors fall detection research in genuine falls rather than safety-constrained simulations. The literature repeatedly frames it as a rare resource for studying whether algorithms remain effective under realistic conditions of long-duration monitoring, participant-wise generalization, class imbalance, and heterogeneous sensors (Aderinola et al., 15 Sep 2025). In the broader benchmarking context, it functions as a clinically oriented “wild” domain that is conceptually different from both staged laboratory datasets and open-web accident videos (Schneider et al., 26 May 2025).

Several limitations are also explicit. First, sensor heterogeneity is inherent to the repository: placements differ, sensor combinations differ, and sampling rates differ (Aderinola et al., 15 Sep 2025). This is why recent work narrows analysis to L5 and 100 Hz. Second, the recent streaming framework uses accelerometer magnitude only, even though the underlying repository may include gyroscope and magnetometer channels (Aderinola et al., 15 Sep 2025). Third, the dataset is clinically oriented and mainly older-adult, so caution is warranted when generalizing to younger populations or materially different clinical cohorts (Aderinola et al., 15 Sep 2025). Fourth, although FARSEEING is relatively large for real-world fall data, the paper explicitly notes that the lack of public real-world fall datasets remains a major limitation (Aderinola et al., 15 Sep 2025).

The methodological implications are correspondingly clear. Continuous-monitoring evaluation should avoid pre-cut fall windows and participant leakage; windows should include pre-fall, impact, and post-fall phases; and event-level criteria with asymmetric tolerance should be preferred to purely framewise decisions (Aderinola et al., 15 Sep 2025). In the video-benchmark literature, FARSEEING also motivates a staged-to-clinical generalization question parallel to OmniFall’s staged-to-wild evaluation on OOPS-Fall (Schneider et al., 26 May 2025). This suggests that future work integrating FARSEEING with taxonomies such as OmniFall’s would provide a direct measure of how far staged representations transfer into clinical deployment domains.

In that sense, FARSEEING serves two roles simultaneously: as a repository of clinically verified real-world falls and as a methodological corrective against overly optimistic evaluation practices.