In this paper, an improved bounding box regression technique is studied for CNN-based object detection and tracking. The proposed approach uses a CNN-based network to identify and localise objects which are boxed by a region of interest regressor. In many cases, a generated bounding box does not fit tightly to the detected object which leads to inaccurate results for spatial analysis such as geometric warping. In this paper, a technique is designed to improve this situation by means of background modelling, motion analysis, template correlation and Kalman filters. The experimental results suggest that the proposed method outperforms single frame object detector-based tracking techniques with an improvement of over 80% in the context of surveillance camera-based pedestrian analysis.
Paper
Full text
On Improving Bounding Box Regression Towards Accurate Object Detection and Tracking
Semantic Scholar · Computer Science · 2019
Abstract
In this paper, an improved bounding box regression technique is studied for CNN-based object detection and tracking. The proposed approach uses a CNN-based network to identify and localise objects which are boxed by a region of interest regressor. In many cases, a generated bounding box does not fit tightly to the detected object which leads to inaccurate results for spatial analysis such as geometric warping. In this paper, a technique is designed to improve this situation by means of background modelling, motion analysis, template correlation and Kalman filters. The experimental results suggest that the proposed method outperforms single frame object detector-based tracking techniques with an improvement of over 80% in the context of surveillance camera-based pedestrian analysis.