Convolutional Neural Network for Camera Pose Estimation From Object Detections

Abstract

Known scene geometry and camera calibration parameters give important information to video content analysis systems. In this paper, we propose a novel method for camera pose estimation based on people observation in the input video captured by static camera. As opposed to previous techniques, our method can deal with false positive detections and inaccurate localization results. Specifically, the proposed method does not make any assumption about the utilized object detector and takes it as a parameter. Moreover, we do not require a huge labeled dataset of real data and train on the synthetic data only. We apply the proposed technique for camera pose estimation based on head observations. Our experiments show that the algorithm trained on the synthetic dataset generalizes to real data and is robust to false positive detections.

A Distributed Tracking Algorithm for Counting People in Video by Head Detection

“…For the future work we plan to use automatic camera calibration algorithms (see fig. 5) [20,21]. This will allow to perform person tracking on the ground map and further improve accuracy of people counting.…”

Camera parameters estimation from pose detections

“…Therefore we are limited to the comparison to another methods of automatical surveillance camera calibration, of which the most (i.e. [11], [13], [16]) neither have an available implementation nor report more than a few results on the real videos (in some cases the videos themselves are unavailable); an exception is [19]. There also are several traffic surveillance camera calibration methods (i.e.…”

“…Shalnov et al [19] estimate extrinsic camera parameters from human head detections obtained with [18], hence it's possible to test our and their methods on the same data. Fortunately the head detector also proves to be applicable for visualized event-based data.…”

“…The results of evaluation are presented in Table 5. The proposed method outperforms [19] even though the latter gets camera focal length as additional input. The results of our method on event-based data are close to the results on RGB data.…”

“…However the parameters estimated by these methods are insufficient to recover the scene geometry. Shalnov et al [19] use CNN that takes focal length and human head detections as input and estimates camera location and orientation. The latter approach suits well the typical surveillance scenarios, however it requires focal length to be known.…”