2015 Volume 27 Issue 5 Pages 813-825
A human detection method for omnidirectional images with distortion and appearance change is proposed in order to improve detection accuracy without camera calibration. The method is based on deep convolutional neural network, and a training data increase method is also proposed with applying translation, scaling, rotation, and brightness change operators to manually prepared training data. In the evaluation experiment with actual omnidirectional images, under the condition at false positive per detection window 0.001, the proposal achieves miss rate 28.2% and miss rate of a combination of HOG and Real AdaBoost, which is standard method in object detection, is 77.5%. In addition, it is validated that the proposed training data increase method improves detection accuracy, and relationships between visualized feature maps and detection accuracy are discussed. The proposal provides improvement of human detection accuracy without estimating camera parameters, where camera calibration is difficult in situations such as large number of cameras, cameras on distant places, and unknown camera models.
