Abstract: A lane line recognition method, a device and a storage medium are provided. The position information of lane lines is determined by first detecting a current frame image collected by a vehicle and determining a plurality of detection frames where the lane lines in the current frame image are located, determining a connection area according to the position information of the plurality of detection frames where the connection area includes the lane lines, and then performing edge detection on the connection area and determining the position information of the lane lines in the connection area. That is to say, the position information of the lane lines is obtained by first dividing the current frame image into a plurality of detection frames, then connecting the detection frames to obtain the connection area including the lane lines, and then performing edge detection on the connection area.

Abstract: A fatigue detecting method includes: obtaining an eye image of a driver (S101); inputting the eye image into a target convolutional neural network, to obtain eye-state data contained in the eye image, wherein the eye-state data represents a degree of opening of an eye of the driver, and the target convolutional neural network is obtained by training a convolutional neural network by using eye-image samples that are collected in advance (S102); and in a condition that respective eye-state data of a plurality of frames of eye images are obtained, according to the respective eye-state data of the plurality of frames of eye images, determining whether the driver is in a fatigue state (S103). Therefore, the method can more easily determine whether the driver is in the fatigue state, and, by using the target convolutional neural network, can realize real-time detection on whether the driver is in the fatigue state.

Abstract: A simulated tracking shot is generated from an image sequence in which a foreground feature moves relative to a background during capturing of the image sequence. The background is artificially blurred in the simulated tracking shot in a spatially-invariant manner corresponding to foreground motion relative to the background during a time span of the image sequence. The foreground feature can be substantially unblurred relative to a reference image selected from the image sequence. A system to generate the simulated tracking shot can be configured to derive spatially invariant blur kernels for a background portion by reconstructing or estimating a 3-D space of the captured scene, placing virtual cameras along a foreground trajectory in the 3-D space, and projecting 3-D background points on to the virtual cameras.

Abstract: Joint video deblurring and stabilization techniques are described. In one or more implementations, a deblurring and stabilization module is configured to jointly deblur and stabilize a video by grouping video frames into spatial-neighboring frame clusters, and building local mesh homographies for video frames in each spatial-neighboring frame cluster.

Abstract: Joint video deblurring and stabilization techniques are described. In one or more implementations, a deblurring and stabilization module is configured to jointly deblur and stabilize a video by grouping video frames into spatial-neighboring frame clusters, and building local mesh homographies for video frames in each spatial-neighboring frame cluster.

Abstract: A simulated tracking shot is generated from an image sequence in which a foreground feature moves relative to a background during capturing of the image sequence. The background is artificially blurred in the simulated tracking shot in a spatially-invariant manner corresponding to foreground motion relative to the background during a time span of the image sequence. The foreground feature can be substantially unblurred relative to a reference image selected from the image sequence. A system to generate the simulated tracking shot can be configured to derive spatially invariant blur kernels for a background portion by reconstructing or estimating a 3-D space of the captured scene, placing virtual cameras along a foreground trajectory in the 3-D space, and projecting 3-D background points on to the virtual cameras.