KJCKorea
Journal Central

Recently, 360° panorama generation using multiple camera images has been widely adopted in autonomous driving and video surveillance systems. However, when detailed observation of a specific region is required, zooming into a low-resolution panorama generated for real-time processing significantly degrades image quality. This paper proposes an inverse homography-based selective high-resolution Region of Interest (ROI) reconstruction method. The proposed method tracks a user-selected ROI via inverse homography to choose two original 4K images, and reuses the low-resolution homography by applying only a scale transformation. The high-resolution ROI is reconstructed by warping only the coordinates of the final crop region when stitching the two images, a process performed only upon user selection so that it does not burden the real-time main pipeline. Experimental results show that the proposed method reduces computational cost while improving the visual quality of zoomed regions and maintaining real-time performance.

This study proposes an SVM+PUK-LSTM hybrid model for EEG-based emotion recognition and evaluation of generalization to unseen subjects. Raw EEG signals from the DEAP dataset were resampled to 256Hz and processed into 3-s non-overlapping epochs, from which a 160-dimensional PSD feature vector was constructed using 32 EEG channels and five frequency bands; emotional states were defined into four Arousal–Valence-based classes: HAHV, HALV, LAHV, and LALV. The proposed model achieved an Accuracy of 98.42±3.82% and a Macro F1-score of 97.55±6.31% under 10-fold cross-validation, and an Accuracy of 90.11±7.91% and a Macro F1-score of 88.38±11.04% under 32-fold subject-independent LOSO evaluation. Under the same LOSO protocol and random seed of 1, comparisons with standalone LSTM and SVM+RBF-LSTM showed statistically significant differences in Accuracy using the Wilcoxon signed-rank test with Holm–Bonferroni correction. These results demonstrate the proposed model's 4-class EEG emotion classification performance and potential for generalization to unseen subjects.

In this paper, the Failure-State Assessment and Transition Estimation Scheme(FATES) for Tracking-by-Detection is proposed. FATES transforms degradation in MOT metrics into multi-label failure compositions and perturbation-space transition intervals using four labels: recall loss, fragmentation, identity-association failure, and localization instability. Rules selected from development data are evaluated against predefined operational event references using 1,260 synthetic runs and 45 window–tracker samples from MOT17, MOT20, and DanceTrack. On synthetic data, agreement with these references yielded 96.8% Macro-F1, 96.9% Micro-F1, and 91.1% exact match. External evaluation was limited to supported failure states and did not assess transitions. An offline application improved mean IDF1 from 64.0% to 67.8% and MOTA from 53.9% to 69.9% on 630 synthetic outputs. On 45 external samples, the gains were limited: IDF1 increased from 55.4% to 55.6% and MOTA from 50.9% to 51.3%.