Multi-part Detection Of Human Based On Convolutional Neural Network

Human detection aims to detect human in pictures and offer the location information of them. It's widely used in motion analysis, intelligent monitoring system, driving assistance system, et al. And it is one research hotspot of object detection. As an object detection algorithm which is based on regions, R-CNN combines selective search, convolutional neural network, support vector machine and non-maximum suppression. This method achieves good performance on human detection. But just as other detection methods which are based on the whole bodies, it cannot offer the location of human's parts and performs not so good when faced with flexible human bodies and occlusion. Compared to detection methods based on the whole bodies, detection methods based on the parts can handle these questions.We propose a multi-part detection algorithm of human based on convolutional neural network on the basis of R-CNN. In this algorithm, we train an R-CNN model based on human's multi-parts and put spatial geometric constraints on it. This algorithm not only achieves good performance on human detection, but also offers the location of human's multi-parts explicitly. In this algorithm, we first generate an R-CNN model based on human's multi-parts. After manually labeling a data set that can offer the location of human's multi-parts, we train that model with the data set. For detection of that model, we can get region proposal bounding boxes from each picture, each region proposal bounding box will get a SVM score which is based on its convolutional neural network features, and we use non-maximum suppression to kick out region proposal bounding boxes that overlap too much with others. In this algorithm, we secondly put spatial geometric constraints on region proposal bounding boxes, and then we will get prediction bounding boxes groups. The details are as follows: Each region proposal bounding box will be assumed as a human's whole body and other region proposal bounding boxes which lead to the highest product of geometric constraints scores and SVM scores will be assumed as the human's parts. In this way, we can get region proposal bounding boxes groups. We get prediction bounding boxes groups by choosing region proposal bounding boxes groups that have higher score than the threshold. Geometric constraints include spatial position constraints, Gaussian Mixture Model constraints and K-nearest Neighbors constraints. The experimental results show that the method in this paper can achieve a good performance towards human detection, and it is of great practical value.