Multiple-Kernel Based Vehicle Tracking Using 3D Deformable Model and Camera Self-Calibration.

摘要: Tracking of multiple objects is an important application in AI City geared towards solving salient problems related to safety and congestion urban environment. Frequent occlusion traffic surveillance has been a major problem this research field. In challenge, we propose model-based vehicle localization method, which builds kernel at each patch the 3D deformable model associates them with constraints space. The proposed method utilizes shape fitness evaluation besides color information track robustly efficiently. To build car models fully unsupervised manner, also implement evolutionary camera self-calibration from tracking walking humans automatically compute parameters. Additionally, segmented foreground masks are crucial modeling adaptively refined by multiple-kernel feedback tracking. For object detection/classification, state-of-the-art single shot multibox detector (SSD) adopted train test on NVIDIA Dataset. improve accuracy categories only few objects, like bus, bicycle motorcycle, employ pretrained YOLO9000 multi-scale testing. We combine results SSD based ensemble learning. Experiments show that our system outperforms both segmentation detection.