PaperlayerView exact revision on arXiv

Source-linked AI summary

Beyond Gait: Person Identification from Millimeter-Wave Point Clouds Across Activities of Daily Living

Xilai Wang, Zixiong Han, Saad Rhanmouni, Chenzhe Zhao, Yunze Lu, Miodrag Bolic

arXiv:2609.08818v1cs.CV

TL;DR

Evidence on mmWave identification across heterogeneous daily activities remains limited, while indoor walking may be brief and interrupted. This paper tests activity-conditioned identity experts and reports improved identification outcomes, including ReID benefits beyond activity-specific gallery restriction.

  • Problem

    Evidence on mmWave identification across heterogeneous daily activities remains limited, and ADL sequences contain varied states and transitions that a shared identity model must accommodate.

    Source: §I · p. 1

  • Method

    The paper uses a human activity recognition router to dispatch each mmWave clip to an activity-specific identity expert.

    Source: §I · p. 1

  • Results

    The HAR-routed framework achieves higher mean ID accuracy than evaluated shared-model alternatives, and its ReID benefit extends beyond restricting retrieval to an activity-specific gallery.

    Source: 3) Routing Supervision · p. 8, §VIII · p. 10

  • Takeaways & Limitations

    The results support activity conditioning as a useful approach to extracting identity information from heterogeneous ADLs under controlled collection and evaluation protocols.

    Source: §VIII · p. 10

    The fixed sensor geometry and scripted protocol leave unestablished whether routing gains hold for uncontrolled behavior, different rooms, or different radar positions.

    Source: §VII · p. 9

Abstract

from arXiv · show

Person identification from millimeter-wave (mmWave) point clouds has mainly relied on gait. Indoor walking, however, is often brief and interrupted, while other activities of daily living (ADLs) may provide complementary identity information. We investigate identification across seven ADLs using mm-ADL, a new point-cloud dataset collected from 11 subjects under a controlled protocol. This extension introduces heterogeneous states and transitions whose spatial and temporal characteristics vary with activity. We therefore study whether activity can provide useful context for learning identity representations. We propose an activity-conditioned framework in which a human activity recognition router dispatches each clip to an activity-specific identity expert. The framework is implemented as a supervised mixture of experts, using a dual-stream static-dynamic PointNet (DS-SDPNet) to combine time-aggregated spatial structure with frame-to-frame information. We evaluate closed-set identification (ID) and subject-disjoint re-identification (ReID). With learned hard routing, ID accuracy increases from 62.1% to 68.0%. In a two-occupant ReID setting, hard routing increases mAP from 57.2% to 75.4% and Rank-1 accuracy from 59.1% to 82.1%. Under a matched gallery partition, activity-specific experts also outperform a shared embedding, showing that the gain extends beyond restricting the gallery. These results support the feasibility of using ADLs beyond gait for identification and the value of activity conditioning under controlled indoor conditions.

I. INTRODUCTION · II. RELATED WORK

The introduction motivates identifying people from mmWave point clouds during activities beyond gait and proposes activity-conditioned identity experts with a dual-stream backbone. The framework is evaluated on seven ADLs in a controlled 11-subject dataset.

I. INTRODUCTION

  • Gait-based identification depends on informative walking sequences, which may be brief and interrupted indoors, while ADLs such as sitting, lying down, and getting up offer additional observations.Identification across heterogeneous daily activities remains less explored than gait recognition in public mmWave point-cloud datasets.Source: §I · p. 1
  • ADL sequences contain heterogeneous states and transitions that change both body configuration and temporal evolution, challenging a monolithic identity model.The same vertical extent, for example, has different interpretations for a standing person and a person lying down.Source: §I · p. 1
  • The proposed framework uses an independently trained activity-recognition router to dispatch clips to activity-specific identity experts, with hard top-1 or soft probability-weighted routing.This supervised mixture of experts tests whether learning identity distinctions within each activity is more effective than using one shared model.Source: §I · p. 1
  • DS-SDPNet combines spatial point distributions aggregated across a clip with frame-to-frame evolution as the identity-expert backbone.The introduction also describes comparisons on mm-ADL and public mmGait, plus component ablations, to assess the backbone.Source: §I · pp. 1–2
  • The study evaluates seven ADLs from 11 subjects in the controlled mm-ADL dataset using closed-set identification and subject-disjoint re-identification.The ReID evaluation considers a two-occupant scenario with held-out identities matched against an enrolled gallery.Source: §I · pp. 1–2
  • Comparisons with monolithic, joint multi-task, and end-to-end MoE models examine activity integration, while matched-gallery analysis tests the contribution of activity-specific embeddings.The introduction frames these analyses as examining explicit activity conditioning, learned-routing cost, and gallery partitioning in ReID.Source: §I · p. 2

A. Activity Biometrics … A. Point-Cloud Generation

Prior work motivates using both activity content and personal movement style for identity, while this study tests activity-conditioned identity learning for heterogeneous mmWave ADLs. Its point-cloud representation is sparse and can lose fully stationary subjects, motivating temporal modeling.

A. Activity Biometrics

  • Body configuration and personal activity style have supported identification across camera, RGB/RGBD, RFID, and wearable-inertial sensing –.These studies motivate treating more than gait as a potential source of identity information.Source: A. Activity Biometrics · p. 2
  • Radar biometrics have used motion signatures, gestures, and gait, motivating extension of mmWave point-cloud identification to whole-body ADL states and transitions, –.Prior work includes UWB and Doppler radar motions, mmWave gesture-and-identity recognition, and gait identification and ReID.Source: A. Activity Biometrics · p. 2

B. Integrating Activity and Identity

  • Activity information can supervise a shared activity-and-identity representation or select a corresponding identity model,,,,,,.The paper frames these as distinct approaches for using activity content alongside personal movement style.Source: B. Integrating Activity and Identity · p. 2
  • The framework routes clips by HAR activity to identity experts and evaluates both ID classification and ReID metric learning, including activity-restricted galleries.The evaluation covers training identities and retrieval for held-out identities.Source: B. Integrating Activity and Identity · p. 2
  • Unlike jointly learned sparse MoE gates and experts [21], this supervised MoE uses activity labels to specialize experts and trains the HAR router independently.The compared training designs test whether explicit activity conditioning is useful in this setting.Source: B. Integrating Activity and Identity · p. 2

A. Point-Cloud Generation

  • FMCW processing estimates range, radial velocity, azimuth, and elevation, then detects and tracks points represented as [x, y, z, vr, SNR].The pipeline applies range FFT, static-clutter suppression, Capon beamforming, CFAR detection, and group tracking.Source: A. Point-Cloud Generation · p. 2, Figure 1 · p. 3
  • Sparse, unstable point clouds with tens of points per subject per frame lack color or texture, and static-clutter removal suppresses fully stationary subjects.A single frame cannot describe body configuration; near-static activities are mainly observable through residual movement before a stationary posture, motivating temporal-window modeling and a static–dynamic architecture.Source: A. Point-Cloud Generation · pp. 2–3

B. ADL Taxonomy … B. Point-Cloud Preprocessing

The paper defines seven observable ADLs as states and transitions, then uses an activity router to dispatch normalized point-cloud clips to identity experts. Preprocessing centers each frame horizontally while preserving height as a body-geometry cue.

B. ADL Taxonomy

  • The seven-activity taxonomy comprises three states—Standing/Walking, Sitting, and Lying Down—and four transitions: Stand-to-Sit, Sit-to-Lie, Lie-to-Sit, and Sit-to-Stand.The taxonomy follows a state-and-transition approach to describing consecutive indoor motion.Source: B. ADL Taxonomy · p. 3, Figure 2 · p. 3
  • Standing and walking are combined because indoor movement often includes brief walks, starts, stops, and halts; sitting and lying are treated as near-static states with residual postural motion.The taxonomy describes observable motion intervals, since a fully stationary subject is largely removed with the background and does not provide a stable point cloud.Source: B. ADL Taxonomy · p. 3, Figure 3 · p. 3

A. Framework Overview

  • The framework assigns activity classification to the router and identity representation to activity-specific DS-SDPNet experts, with activity probabilities controlling expert processing.The normalized clip is processed by the HAR router before dispatch to the experts.Source: A. Framework Overview · p. 3, Figure 4 · p. 4
  • The same activity-routed framework supports both an ID classification head and a ReID metric-learning head.Source: A. Framework Overview · p. 3, Figure 4 · p. 4

B. Point-Cloud Preprocessing

  • Each clip is represented as P ∈ RT×N×C with C = 4 features (x, y, z, vr), and each frame is centered in the horizontal plane to remove absolute location.The horizontal coordinates describe where a person stands rather than how the person moves.Source: B. Point-Cloud Preprocessing · p. 3
  • Preprocessing leaves height z unchanged as a ground-relative body-geometry cue, and the normalized clip is shared by the router and experts.Source: B. Point-Cloud Preprocessing · p. 3

C. DS-SDPNet Identity Expert … D. HAR Router and Routing Strategies

DS-SDPNet combines frame-to-frame dynamics with the accumulated spatial shape of a clip for identity representation. A PointNet+LSTM HAR router dispatches clips to activity-specific experts using hard or soft routing, with an oracle router used to assess learned routing.

C. DS-SDPNet Identity Expert

  • The final DS-SDPNet identity feature fuses separately extracted motion and accumulated-cloud shape representations before passing them to an ID or ReID head.The dynamic stream captures how an activity is performed, while the static stream reflects stature and build.Source: C. DS-SDPNet Identity Expert · p. 4, Figure 5 · p. 5

1) Dynamic Stream:

  • The dynamic stream encodes each frame with a shared-weight PointNet, then pools bidirectional LSTM outputs across time into a 4h-D feature.Average pooling captures persistent trajectory components, while maximum pooling retains strong short-lived responses.Source: 2) Static Stream and Feature Fusion · p. 4

2) Static Stream and Feature Fusion:

  • The static stream treats all T frames as one T ×N-point cloud and uses a separate PointNet with global max pooling to produce a 256-D feature.Discarding temporal order yields a denser spatial envelope than a single sparse radar frame.Source: 2) Static Stream and Feature Fusion · p. 4
  • The 256-D static feature and 4h-D dynamic feature are concatenated and batch normalized into fused feature z.The fused representation is 512-D for ID with h = 64 and 768-D for ReID with h = 128.Source: 2) Static Stream and Feature Fusion · p. 4

D. HAR Router and Routing Strategies

  • The HAR router processes frame-wise PointNet descriptors through FC–ReLU–dropout, a bidirectional LSTM, and a second FC–ReLU–dropout block.The LSTM has 128 hidden units per direction; both dropout blocks use probability 0.1.Source: D. HAR Router and Routing Strategies · p. 4
  • The router concatenates ordered projected time steps into a 64T-D vector and classifies them into seven activity logits.For T = 20, the concatenated vector is 1280-D; retaining order distinguishes opposite transitions such as Stand-to-Sit and Sit-to-Stand.Source: D. HAR Router and Routing Strategies · p. 4
  • Hard routing selects the most probable expert, while soft routing combines the top three experts using router probabilities.For ID, expert class distributions can be combined directly; for ReID, weighted embeddings are L2 normalized before matching. A ground-truth-activity oracle router measures the upper bound and isolates learned-routing cost.Source: D. HAR Router and Routing Strategies · pp. 4–5

E. ID and ReID Heads · 1) Identification Head:

The ID and ReID tasks share a DS-SDPNet backbone but use different heads and training losses, with each activity expert holding its own copy of the selected head. The identification head classifies clips among training subjects using a multilayer classifier and mini-batch cross-entropy.

E. ID and ReID Heads

  • Each activity expert has its own copy of the selected head, while ID and ReID share a DS-SDPNet backbone but use different heads and training losses.Source: E. ID and ReID Heads · p. 5

1) Identification Head:

  • Closed-set ID assigns each clip to one of the K subjects seen during training.Source: 1) Identification Head · p. 5
  • The identification head maps 256 + 4h inputs through 256- and 128-unit fully connected layers before a K-class classifier.The two hidden layers each use batch normalization, ReLU, and dropout.Source: 1) Identification Head · p. 5
  • The subject classifier is trained with mini-batch cross-entropy on subject logits and ground-truth labels.The passage defines the logits per clip but does not provide the loss formula in the supplied text.Source: 1) Identification Head · p. 5
  • The HAR router uses corresponding activity cross-entropy, and expert losses do not update the router.Independent router and expert training keeps activity semantics explicit and allows one router to serve different ID and ReID expert banks.Source: 1) Identification Head · p. 5

2) Re-identification Head: · V. EXPERIMENTAL SETUP · A. mm-ADL Dataset

The section describes a 256-dimensional metric-learning head for open-set re-identification and the controlled collection and processing of mm-ADL. The dataset contains 3,850 clips across 11 subjects and three activity protocols.

2) Re-identification Head:

  • The open-set ReID head maps fused DS-SDPNet features through two fully connected layers to a 256-dimensional embedding, then applies L2 normalization.Its first layer maps the (256 + 4h)-dimensional feature to 512 dimensions before producing the embedding.Source: 2) Re-identification Head · p. 5
  • Batch-hard mining selects each anchor’s farthest positive and nearest negative, and the margin loss pulls same-subject clips together while separating different subjects.Anchors without a valid positive in the mini-batch are excluded; the margin is 0.2.Source: 2) Re-identification Head · p. 6

A. mm-ADL Dataset

  • Eleven subjects performed walk, sit, and lie sequences in a 4 × 8 m room, using an IWR6843ISK-ODS radar and synchronized RGB ground truth.The radar had a 120° field of view, was mounted at approximately 1.5 m, and monitored 0.9–7 m.Source: A. mm-ADL Dataset · p. 6
  • Each subject repeated each activity sequence 50 times under fixed radar geometry; collection spanned multiple days, varied outerwear, and used synchronized video for activity annotations.Sequences included random walking with turns and halts, sitting and standing, and lying down and getting up; the study received University of Ottawa REB approval H-11-25-12304.Source: A. mm-ADL Dataset · p. 6
  • Activity durations vary: transitions are short, while stationary sitting and lying intervals are limited because clutter removal suppresses point clouds after subjects become fully still.Those intervals therefore retain near-static states with residual postural motion.Source: A. mm-ADL Dataset · p. 6
  • Sequences were segmented into 20-frame windows labeled by their longest-duration activity, with 128 points per frame and 50 non-overlapping clips retained per activity per subject.Frames were randomly downsampled or padded by resampling available points, following.Source: A. mm-ADL Dataset · p. 6
  • The resulting mm-ADL dataset contains 3,850 clips and 77,000 frames, each represented as P ∈ R20×128×4.The dataset is publicly available at https://github.com/OwenHan10/mm-ADL.git.Source: A. mm-ADL Dataset · p. 6

B. Evaluation Protocols · C. Implementation Details · D. Evaluation Metrics

The evaluation uses separate closed-set identification and subject-disjoint ReID protocols, with accuracy for ID and HAR and ranking metrics for ReID. The sections also specify model-training settings and how ReID ranking quality is assessed.

B. Evaluation Protocols

  • Closed-set ID uses all 11 subjects in training and testing, with five-fold contiguous clip partitions yielding 3,080 training and 770 test clips per fold.The router and experts are trained from scratch in each fold, and test clips are scored even when misrouted.Source: B. Evaluation Protocols · p. 6
  • ReID uses six folds with nine training subjects and two held-out subjects; each held-out subject contributes 40 gallery clips and 10 query clips per activity.The protocol models a two-occupant indoor scenario with a fixed enrolled gallery, not a large-population ReID benchmark.Source: B. Evaluation Protocols · p. 6

C. Implementation Details

  • Models train for 500 epochs with Adam, weight decay 10^-3, dropout 0.1, and a fixed random seed; ID and ReID experts use different learning rates, batch sizes, and hidden dimensions.ID experts use learning rate 10^-3, batch size 25, and h = 64; ReID experts use learning rate 10^-4, batch size 60, h = 128, and a 256-D embedding. The HAR router uses learning rate 10^-3 and batch size 25; soft routing retains k = 3 experts with τ = 1.Source: C. Implementation Details · p. 6

D. Evaluation Metrics

  • ID and HAR performance is measured by accuracy, the proportion of test clips whose predicted subject or activity label matches ground truth, averaged across folds.Source: D. Evaluation Metrics · p. 6
  • ReID reports Rank-1 (CMC@1), which measures whether the first returned identity is correct, with results averaged across folds.Source: D. Evaluation Metrics · p. 7
  • ReID also reports mAP, which measures how consistently all correct clips rank ahead of impostors.Source: D. Evaluation Metrics · p. 7

VI. RESULTS AND ANALYSIS … 1) HAR Router:

Oracle-routed results show identity information across all seven mm-ADLs, while a separate leave-one-subject-out evaluation finds that activity can be estimated accurately for held-out subjects. The analysis also examines learned activity conditioning, ReID, backbones, ablations, and feature visualizations.

VI. RESULTS AND ANALYSIS

  • The analysis proceeds from individual-ADL identity information to activity conditioning and subject-disjoint ReID, then examines gallery partitioning, embeddings, backbones, ablations, and feature visualizations.It separates the effects of gallery partitioning and activity-specific embeddings in its ReID analysis.Source: §VI · p. 7

A. ADL Identification Feasibility

  • 80.8% accuracy was achieved for Sit-to-Lie, and all seven ADLs exceeded 62% oracle-routed ID accuracy across 11 subjects.The evaluation used ground-truth activity labels to select the corresponding identity expert.Source: A. ADL Identification Feasibility · p. 7, Table III · p. 7
  • Transitions generally outperformed state activities, while Standing/Walking had the lowest accuracy in the oracle-routed evaluation.Standing/Walking combines brief walking with starts, stops, halts, and standing.Source: A. ADL Identification Feasibility · p. 7
  • Differences in body configuration and temporal information may explain variation across ADLs, but these results do not isolate geometry from movement-pattern contributions.Sitting and Lying Down include residual postural motion before subjects become fully stationary.Source: A. ADL Identification Feasibility · p. 7
  • The oracle-routed findings support identity recognition from non-gait ADLs under the controlled protocol when activity is known.The passage notes that the table also includes three mmWave point-cloud identity baselines, whose backbone comparison is discussed separately.Source: A. ADL Identification Feasibility · p. 7

1) HAR Router:

  • 95.7% overall accuracy was achieved by PointNet+LSTM in leave-one-subject-out HAR, with every activity at or above 94.0%.The evaluation used the same input and protocol for the compared HAR methods.Source: 1) HAR Router · p. 7, Table IV · p. 7
  • PointNet+LSTM had higher overall mean accuracy than DVCNN and m-Activity, indicating that activity could be estimated for held-out subjects.The ID and ReID pipelines evaluate routing cost separately by comparing learned and oracle routing.Source: 1) HAR Router · p. 7

2) Activity–Identity Integration: … 2) Gallery Partition and Expert Specialization:

Activity-conditioned identity models improve identification and subject-disjoint retrieval in the evaluated protocols. Matched-gallery results show that ReID gains extend beyond restricting the gallery to include activity-specific identity experts.

1) Learned Routing and Oracle Reference:

  • 75.4% mAP and 82.1% Rank-1 under hard routing exceed the monolithic model’s 57.2% and 59.1%, respectively, in two-occupant ReID.The learned-routing results are close to oracle routing’s 76.2% mAP and 84.1% Rank-1, indicating limited performance cost from using the learned router in this protocol.Source: Table V · p. 7, 1) Learned Routing and Oracle Reference · p. 8

2) Activity–Identity Integration:

  • 68.0% ID accuracy with HAR hard routing and 69.0% with soft routing exceed 64.3% for joint multi-task learning and 62.1% for the monolithic model, near the 70.2% oracle result.The small numerical advantage of soft over hard routing does not establish a statistically significant difference.Source: Table V · p. 7, 2) Activity–Identity Integration · p. 7

3) Routing Supervision:

  • 56.0% ID accuracy for the end-to-end MoE variant based on [21] is below both HAR-routed variants under the evaluated training designs.The seven-expert model and capacity-matched gate are trained jointly from the identity objective without activity labels.Source: 3) Routing Supervision · pp. 7–8
  • Gate diagnostics show 0.600 ± 0.018 ID accuracy with dense softmax and 0.513 ± 0.054 with hard top-1 routing, alongside partial activity alignment and uneven expert use.The gate has 0.280 ± 0.155 activity NMI, 0.363 ± 0.130 activity purity, and utilization entropy 1.261 ± 0.646 versus the uniform-use maximum ln 7 = 1.946, or roughly 3.5 effective experts. The 0.837 ± 0.025 any-expert-correct diagnostic uses true identity and is not an available inference rule; the gate is trained for identity prediction, while HAR-routed accuracy supports explicit activity conditioning.Source: 3) Routing Supervision · p. 8

C. Subject-Disjoint Re-identification

  • The subject-disjoint protocol uses two held-out subjects per fold, embedding enrolled-gallery and query clips without retraining and matching each query within its selected expert’s gallery.This evaluates activity-conditioned retrieval for two enrolled identities.Source: C. Subject-Disjoint Re-identification · p. 8

2) Gallery Partition and Expert Specialization:

  • Under a matched evaluation protocol, approximately three quarters of the total mAP improvement occurs when the activity-specific expert bank replaces the shared model, with the remaining quarter from restricting the gallery.Source: Table VI · p. 8, 2) Gallery Partition and Expert… · p. 8

3) Routing Error: … 2) Component Ablations:

The analysis finds DS-SDPNet strongest overall among the compared backbones and supports combining its static and dynamic streams. It also favors hard top-1 routing for the present ReID implementation, while finding little gallery impact from routing contaminants.

3) Routing Error:

  • Removing mis-routed gallery clips changes mAP by only +0.0040 and Rank-1 by −0.0057, so gallery contamination does not meaningfully degrade results.This gallery analysis is distinct from the learned-versus-oracle comparison of routing errors’ overall effect on the pipeline.Source: 3) Routing Error · p. 8

4) Soft Routing:

  • Soft routing yields 81.8% Rank-1 but 55.4% mAP, below the monolithic mean, supporting hard top-1 routing for the present ReID implementation.Independently trained ReID experts lack aligned metric spaces, so weighted embedding combinations can disturb ranking; ID expert class distributions share a subject-class axis.Source: 4) Soft Routing · p. 8

D. Backbone and Component Analysis — 1) Identity Backbone:

  • The reproduced point-cloud identity baselines are SRPNet, HDNet, and DSFE+LGTE, models developed mainly for gait.They provide the closest point-cloud identity baselines for evaluating ADL identification under the same input and folds.Source: 1) Identity Backbone · p. 8
  • On single-person mmGait, DS-SDPNet reaches 81.1% ID accuracy, numerically close to HDNet at 80.8%, without establishing a significant difference.The comparison uses the same 20-frame window and five-fold protocol.Source: 1) Identity Backbone · p. 8
  • DS-SDPNet’s 1.24 G FLOPs exceed SRPNet and DSFE+LGTE but are substantially below HDNet’s 3.49 G FLOPs.This comparison concerns the identity backbone.Source: 1) Identity Backbone · p. 8
  • DS-SDPNet has the highest overall mean ID, ReID mAP, and ReID Rank-1 among the compared backbones on mm-ADL.It also has the highest mean ID accuracy on six of seven activities.Source: 1) Identity Backbone · p. 8, Table VII · p. 9

D. Backbone and Component Analysis — 2) Component Ablations:

  • For HAR, PointNet-only reaches 92.3% accuracy, LSTM-only 85.6%, and the combined PointNet+LSTM 95.7%.The results support combining frame-wise spatial features with temporal modeling for activity recognition.Source: 2) Component Ablations · p. 9, Table VIII · p. 9
  • For identity, the dynamic-stream variant reaches 69.6% accuracy versus 67.1% for the static stream, while combining both gives the highest mean accuracy at 70.2%.Both streams use spatial coordinates and radial velocity; the static stream aggregates observations across the clip, so these ablations do not isolate body geometry from personal movement style.Source: 2) Component Ablations · p. 9, Table VIII · p. 9

E. Feature Visualization · VII. LIMITATIONS AND FUTURE WORK · VIII. CONCLUSION

The study finds usable identity information across seven ADLs and reports benefits from activity-conditioned identification, while noting that its controlled evaluation does not establish performance in more varied settings. Feature visualizations provide qualitative context; broader validation remains necessary.

E. Feature Visualization

  • Held-out ID visualizations show interspersed subject features for monolithic DS-SDPNet, while Fig. 7 displays its identity-colored embedding.The passage describes this as a qualitative view of the shared model's learned representation.Source: E. Feature Visualization · p. 9, Figure 7 · p. 9
  • Fig. 8 shows activity-separated HAR-router groups and identity groups within activity-specific expert features, but t-SNE projections are supporting observations rather than quantitative evidence.The passage identifies Tables V and VI as the quantitative evidence for activity conditioning.Source: E. Feature Visualization · p. 9, Figure 8 · p. 10

VII. LIMITATIONS AND FUTURE WORK

  • The seven-activity ethogram is a coarse model of continuous daily behavior, and mm-ADL contains 11 subjects with 50 clips per activity.ReID evaluates two enrolled identities with 40 gallery clips per activity per subject; the study calls for a larger dataset to assess more identities and smaller galleries.Source: §VII · p. 9
  • All subjects followed one execution protocol under fixed sensor geometry, so routing gains do not establish benefits for uncontrolled behavior, different rooms, or different radar positions.The scripted protocol may favor activity-specific learning by reducing within-activity variation; future data should include these sources of variation.Source: §VII · p. 9
  • The component ablations do not separate body-geometry from movement-related identity information, leaving the relative contribution of stature and personal movement style unresolved.The results establish usable identity information in recorded ADLs without attributing it primarily to either source.Source: §VII · p. 9
  • Hard routing activates one of seven stored activity experts per clip, motivating future investigation of shared or compressed experts that retain the semantic partition.Learned and oracle routing are close for the evaluated models, but substantial identification errors remain even with oracle activity labels.Source: §VII · p. 9
  • Long-term indoor monitoring needs further validation and temporal context such as location and identity continuity, while larger, more diverse data could support more occupants and open-set rejection.The passage proposes temporal context to correct isolated mis-identifications.Source: §VII · p. 9

VIII. CONCLUSION

  • Activity-conditioned experts achieve higher mean ID accuracy than evaluated shared-model alternatives, support two-occupant subject-disjoint ReID, and retain a matched-gallery benefit beyond restricting retrieval.DS-SDPNet combines time-aggregated spatial structure with frame-to-frame information; broader deployment requires evaluation with more subjects, environments, and uncontrolled behavior.Source: §VIII · pp. 9–10

Related papers

Loading 2609.08818v1…