Person Location And Key Point Detection With Deep Learning

Person location detection and keypoint detection are very challenging tasks.In recent works,methods based on deep convolutional neural networks have achieved greater progress than traditional methods.However,existing methods still have some problems.In this work,we explore the three tasks comprehensively.The main contri-butions are:(1)We extend the general object detection framework Faster R-CNN to do pedes-trian detection.After introducing multi-scale features,we solve the problem of being insensitive to detecting small-sized human body.The average precision of this method outperforms the best MS-CNN at the time on KITTI dataset and the miss rate of this method is lower than the best RPN+BF at the time on Caltech dataset.(2)We improve the latest Stacked Hourglass model by strengthening the receptive field of network neurons and trying a multi-scale fusion.It solves the problem of being insensitive to scale variations.Our method achieved a higher PCKh score than the best associative embedding model at the time.(3)We introduce the discriminative loss function in instance segmentation to do keypoint detection.At the same time,We also propose a novel person proposal group-ing strategy.This method has been evaluated on the MS COCO dataset,achieving better detection results than the best associative embedding method at the time.