Due to the popularity and difficulty of research in the field of computer vision, pedestrian detection has been widely used in auxiliary driving and traffic monitoring. Traditional feature extraction method for pedestrians in complex environment is difficult to effectively capture the distinct characteristics information. And convolutional neural network which is popular at present has some influence on generalization performance because BP algorithm is easy to fall into local minimum value,and with the increase of the network layer, some significant feature information is decreasing layer by layer. In view of the above problems, this paper proposes a pedestrian detection algorithm combining deep sensing features with kernel extreme learning machines. Firstly, on the basis of CNN structure, the front layer features and the deep layer features are fused in two stages, and then sent to the followup layer for learning, the directed acyclic graph network(DAGnet) network is constructed. Then, the depth feature information is classified by the kernel extreme learning machine with high realtime performance and strong generalization ability, and the parameters are optimized by Kfold cross validation;In the detection phase, graphbased visual saliency(GBVS) saliency detection algorithm is used on the feature map learned by the DAGnet network to quickly mark the pedestrian area in the test image, and then sliding window is used to identify the precise position of the pedestrian in the salient area.The experimental results show that the positive detection rate on the INRIA data set and the Caltech data set is higher than 90%,and the detection speed is improved significantly if the accuracy is guaranteed.