% IMPORTANT: The following is UTF-8 encoded.  This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.

@ARTICLE{Aldenhoven:285257,
      author       = {Aldenhoven, Céline Madeleine and Nissen, Leon and
                      Heinemann, Marie and Dogdu, Cem and Hanke, Alexander and
                      Jonas, Stephan and Reimer, Lara Marie},
      title        = {{R}eal-{T}ime {E}motion {R}ecognition {P}erformance of
                      {M}obile {D}evices: {A} {D}etailed {A}nalysis of {C}amera
                      and {T}rue{D}epth {S}ensors {U}sing {A}pple's {ARK}it.},
      journal      = {Sensors},
      volume       = {26},
      number       = {3},
      issn         = {1424-8220},
      address      = {Basel},
      publisher    = {MDPI},
      reportid     = {DZNE-2026-00199},
      pages        = {1060},
      year         = {2026},
      abstract     = {Facial features hold information about a person's emotions,
                      motor function, or genetic defects. Since most current
                      mobile devices are capable of real-time face detection using
                      cameras and depth sensors, real-time facial analysis can be
                      utilized in several mobile use cases. Understanding the
                      real-time emotion recognition capabilities of device sensors
                      and frameworks is vital for developing new, valid
                      applications. Therefore, we evaluated on-device emotion
                      recognition using Apple's ARKit on an iPhone 14 Pro. A
                      native app elicited 36 blend shape-specific movements and 7
                      discrete emotions from N=31 healthy adults. Per frame,
                      standardized ARKit blend shapes were classified using a
                      prototype-based cosine similarity metric; performance was
                      summarized as accuracy and area under the receiver operating
                      characteristic curves. Cosine similarity achieved an overall
                      accuracy of $68.3\%,$ exceeding the mean of three human
                      raters $(58.9\%;$ +9.4 percentage points, $≈16\%$
                      relative). Per-emotion accuracy was highest for joy, fear,
                      sadness, and surprise, and competitive for anger, disgust,
                      and contempt. AUCs were ≥0.84 for all classes. The method
                      runs in real time on-device using only vector operations,
                      preserving privacy and minimizing compute. These results
                      indicate that a simple, interpretable cosine-similarity
                      classifier over ARKit blend shapes delivers
                      human-comparable, real-time facial emotion recognition on
                      commodity hardware, supporting privacy-preserving mobile
                      applications.},
      keywords     = {Humans / Emotions: physiology / Adult / Male / Female /
                      Mobile Applications / Smartphone / Young Adult / Facial
                      Expression / ARKit (Other) / emotion recognition (Other) /
                      face tracking (Other) / real-time (Other) / sensors (Other)},
      cin          = {AG Schneider},
      ddc          = {620},
      cid          = {I:(DE-2719)1011305},
      pnm          = {353 - Clinical and Health Care Research (POF4-353)},
      pid          = {G:(DE-HGF)POF4-353},
      typ          = {PUB:(DE-HGF)16},
      pubmed       = {pmid:41682575},
      pmc          = {pmc:PMC12899966},
      doi          = {10.3390/s26031060},
      url          = {https://pub.dzne.de/record/285257},
}