معماري برخط تشخيص و رديابي چندشئ «سر انسان» با تأكيد بر مقياس‌پذيري و پردازش چندجرياني در DeepStream

Real-Time Architecture for Multi-Object Human Head Detection an‎d Tracking with Emphasis on Scalability an‎d Multi-Stream Processing in DeepStream

مهندسي برق

Abstract Industrial video surveillance an‎d people-flow analytics require stable identity tracking of many individuals across tens of cameras in real time an‎d with low latency. Face based tracking in such settings often suffers from severe appearance changes due to pose, illumination, expression an‎d occlusion, which leads to frequent ID switches an‎d degraded performance. This thesis proposes a practical multi object tracking approach for multi camera environments in which the “head” is used as the primary tracking target an‎d face recognition is applied in a subsequent stage on more stable crops. We first review state of the art multi object trackers based on Siamese networks an‎d detection based pipelines, together with software an‎d hardware platforms (OpenCV, GStreamer an‎d NVIDIA DeepStream), an‎d, according to the customer’s requirements, selec‎t a detection based architecture built on DeepStream an‎d desktop GPUs. On top of this platform, we design an‎d implement a multi camera pipeline for head detection an‎d tracking an‎d facial keypoint extraction. The core detector is a YOLOv8m Pose model trained on a new dataset of 34,579 images constructed by combining the CrowdHuman an‎d WIDER FACE datasets an‎d providing, for each head, a bounding box an‎d up to five facial keypoints. On this dataset the detector achieves 70.6% mAP50 an‎d 43.1% mAP50 95 for head bounding boxes an‎d 34.3% mAP50 an‎d 32.3% mAP50 95 for facial keypoints. Performance eva‎luation shows that, at 1080p resolution an‎d 15 frames per second, the proposed pipeline can process 12 video streams (about 60 faces) on a single RTX 3060 GPU an‎d up to 55 streams (about 275 faces) on an RTX 3090, while maintaining real time operation. Field tests with industrial partners confirm the stability of identities an‎d the practical usability of the system. Overall, the thesis demonstrates that head based tracking, combined with a DeepStream based pipeline an‎d a tailored training dataset, provides a scalable an‎d deployable solution for industrial multi camera surveillance. Keywords: Multi-Object Tracking, Siamese Networks, YOLOv8m-Pose, Pipeline, DeepStream, Face Recohnition

5 فصل

153079