Forward obstacles detection is one of the key tasks in the perception system of autonomous vehicles. The perception solution differs from the sensors and the detection algorithm, and the vision-based approaches are always popular. In this paper, an embedding fast obstacles detection algorithm is proposed to efficiently detect forward diverse obstacles from the image stream captured by the monocular camera. Specifically, our algorithm contains three components. The first component is an object detection method using convolution neural networks (CNN) for single image. We design a detection network based on shallow residual network, and an adaptive object aspect ratio setting method for training dataset is proposed to improve the accuracy of detection. The second component is a multiple object tracking method based on correlation filter for the adjacent images. Based on precise detection result, we use multiple correlation filters to track multiple objects in every adjacent frame, and a multi-scale tracking region algorithm is applied to improve the tracking accuracy at the same time. The third component is fusing the detection method and tracking method based on parallel processing, which can significantly increase the average processing rate for the image stream or video in embedded platform. Besides, our algorithm is tested on KITTI dataset as well as our own dataset, and the experimental results illustrate that our algorithm has high precision and robustness. Meanwhile, we test our algorithm on a popular embedded platform - NVIDIA Jetson TX1, and the average processing rate is approximately 17 fps, which satisfies the high processing speed requirements of autonomous vehicles.