Résumé : À Kinshasa, la circulation routière réunit des véhicules et des usagers très différents dans des espaces où les occlusions, les déplacements désordonnés et la diversité des points de vue compliquent l’observation automatique. Nous étudions Smart-Trafic, une chaîne expérimentale associant détection visuelle, suivi multi-objet, construction d’indicateurs temporels, prévision à court terme et comparaison de stratégies de signalisation. Un détecteur YOLOv8n, adapté à sept catégories locales, est évalué sur un ensemble de 358 images et 482 annotations, dont 56 images et 67 objets réservés au test. Il obtient une précision de 0,5500, un rappel de 0,3502 et une mAP@0,50:0,95 de 0,1803. L’ablation met en évidence, dans les conditions expérimentales retenues, l’intérêt du pré-entraînement et des augmentations de données. Pour le score de pression calculé à partir des vidéos, la persistance présente la plus faible MAE à une minute ; le LSTM obtient les erreurs absolues les plus faibles à trois et cinq minutes, mais son R² à cinq minutes demeure négatif (−0,2184). Dans le modèle de files agrégées, l’heuristique prédictive est la plus performante ; en microsimulation SUMO/TraCI, PPO réduit par rapport au feu fixe la file moyenne de 44,79 %, l’attente de 79,65 % et l’écart d’équité de 41,94 %, au prix d’une diminution du débit de 11,11 %. Ces résultats éclairent les compromis entre qualité de perception, prévision et décisions de contrôle. Ils établissent une faisabilité expérimentale, et non une efficacité démontrée en situation routière réelle.

Mots-clés : systèmes de transport intelligents ; Kinshasa ; YOLOv8n ; ByteTrack ; prévision de trafic ; apprentissage par renforcement ; PPO ; SUMO ; robustesse expérimentale.

Abstract: Urban traffic in Kinshasa involves heterogeneous road users, frequent occlusions, and limited ground-truth data, which complicate the use of computer vision for traffic monitoring. This study examines Smart-Trafic, an experimental pipeline linking object detection, multi-object tracking, short-term forecasting, and simulated traffic-signal control. A YOLOv8n detector fine-tuned for seven locally relevant classes was evaluated on 358 images containing 482 annotations, with 56 images and 67 instances reserved for testing. Test precision, recall, and mAP@0.50:0.95 were 0.5500, 0.3502, and 0.1803, respectively. Ablation results indicate descriptive benefits from transfer learning, data augmentation, and a 640-pixel input resolution. For a video-derived congestion-pressure proxy, persistence achieved the lowest one-minute mean absolute error; LSTM had the lowest errors at three and five minutes, although its five-minute R² remained negative (−0.2184). A predictive heuristic ranked first in the aggregate-queue analytical environment. In a separate SUMO/TraCI evaluation, PPO reduced mean queues by 44.79%, waiting time by 79.65%, and a fairness-gap indicator by 41.94% relative to fixed-time control, while throughput fell by 11.11%. The main contribution is an integrated, resource-aware experimental framework that makes trade-offs between perception reliability, forecasting, and control explicit. No effectiveness in real-world intersection operations is claimed.

Keywords: heterogeneous traffic; Kinshasa; computer vision; YOLOv8n; ByteTrack; traffic forecasting; PPO; SUMO; resource-constrained machine learning.

Title: Smart-Trafic: Approche intelligente de perception visuelle, prévision à court terme et évaluation du contrôle adaptatif de la circulation à Kinshasa

Author: Christophe Ebaka Bongelo, Mukala Patrick, Nsenge Mpia Héritier, Telesphorus Nsumbu Lukamba, Blaise Kapalala Kapenda, Yves Nsota Olang, Emmanuel Moyiba Mayayi

International Journal of Novel Research in Engineering and Science

ISSN 2394-7349

Vol. 13, Issue 2, September 2026 - February 2027

Page No: 8-23

Novelty Journals

Website: www.noveltyjournals.com

Published Date: 30-October-2026

Abstract: Lunar regolith combines a wide compositional range, a broad and repeatedly reworked particle-size distribution, and depth-dependent shear strength, yet these three properties are almost always reported separately in the literature. This study applies Shannon entropy to two of these properties for five Apollo landing sites (12, 14, 15, 16 and 17): the bulk oxide composition of returned fines, and the dispersion of published particle-size statistics. A normalized compositional entropy and a granulometric variability index are computed for each site and combined into a composite SelenEntropic Resilience Index (SERI). Mare-influenced sites with a comparatively even mix of oxides score highest on compositional entropy, while the anorthosite-rich Apollo 17 soil, dominated by a single oxide group, scores lowest. Apollo 14 stands apart on the granulometric side, showing far greater dispersion in reported grain size than any other site, a pattern traceable to known sampling heterogeneity near Cone Crater ejecta. The resulting SERI values are discussed alongside independently reported lunar geotechnical parameters, bulk density, internal friction angle and cohesion, to indicate where compositional or granulometric disorder may warrant closer geotechnical attention before excavation, foundation work or in-situ resource utilization at a given site. The index is presented as a first-order screening tool built entirely from published NASA and peer-reviewed data, and the additional site-resolved sieve and shear-strength measurements needed to extend it into an operational design parameter are identified.

Keywords: Apollo regolith; compositional entropy; geotechnical stability; granulometry; in-situ resource utilization; lunar soil mechanics; particle-size distribution; Shannon entropy.

Title: SelenEntropic Resilience: An Entropy-Based Assessment of Lunar Soil Stability Using NASA Data

Author: Dr Nikolaos Toumasis

International Journal of Novel Research in Engineering and Science

ISSN 2394-7349

Vol. 13, Issue 2, September 2026 - February 2027

Page No: 1-7

Novelty Journals

Website: www.noveltyjournals.com

Published Date: 02-September-2026