A multi-class pedestrian detection network for distorted pedestrians
Jiao Zhang, Jiangjian Xiao, Chuanhong Zhou, Chengbin Peng · 2018
As we all known, pedestrians will perform so many different shape and pose in distorted visual pictures, such as surveillance camera. So the problem is that how can we detect pedestrians efficiently and accurately with limited resources in distorted field of vision. This paper proposes a multi-class pedestrian detection network for distorted pedestrian, according to the level of distortion, pedestrians in the visual field can be regarded as different kinds of targets. Recent deep learning object detectors have shown excellent performance for general object detection and pedestrian detection. Based on Faster R-CNN neural network, the experiment is consisted of training classifier and test stage. During the training stage, we define the multi-classification layer to classify pedestrian into different level of distorted pedestrian. And the gradient descent velocity is basis of classification. In the test part, we evaluate our detectors on the train station pedestrian dataset. The final results show that the speed and precision of pedestrian detection in distorted field of vision are greatly proved.