Information tracking method based on convolutional neural network

[0022] The invention will be further explained with reference to the following drawings and embodiments.

[0023] An information tracking method based on convolutional neural network, such as Figure 1 Shown in, including the following operation steps:

[0024] Combining correlation filtering and convolution neural network to construct correlation filtering convolution neural network, in which correlation filtering network is a layer of convolution neural network formed by correlation filtering algorithm;

[0025] Building a time flow convolution neural network and a space flow convolution neural network on the basis of the correlation filtering convolution neural network built in step one;

[0026] The correlation filtering convolution neural network, the time flow convolution neural network and the space flow convolution neural network are constructed to form a deep network by using the jumping connection mode.

[0027]The deep network is trained until the three models of correlation filtering convolution neural network, time flow convolution neural network and space flow convolution neural network all converge; Comprises the following steps of: taking a frame of a target position as an input, extracting a training block with the target position as the center, obtaining feature extraction and response mapping, wherein the size of the training block obtained in the first frame is five times of the maximum width and height of the target, adopting a deep convolution network for feature extraction, and simultaneously initializing the parameters of a correlation filtering convolution neural network, a time convolution neural network and a spatial convolution neural network to zero-mean Gaussian distribution, The training block of the first frame passes through the deep convolution network to get the feature maps of the third, fourth and fifth layers, and then the feature maps of these three layers are sent into the deep network connected by the correlation filtering convolution neural network, the time convolution neural network and the space convolution neural network, and the three models are trained at the same time until they converge.

[0028] The obtained time-flow convolution neural network and space-flow convolution neural network respectively extract the image block feature information in the current frame and the motion change feature information set of the target between frames in the multi-segment time series, and select the weight of each identical feature information in the motion change feature information set of the target between frames in the multi-segment time series; Specifically, extracting the feature information of the image blocks in the current frame through the spatial flow convolutional neural network specifically includes: determining the target area in the previous frame of the current frame, sampling N image blocks in the current frame through Gaussian sampling, and taking them as the input of the spatial flow convolutional neural network; Complete the zeroing around n image blocks by Numpy function, and finally output the image features, in which the spatial flow convolution neural network includes three convolution layers and two pooling layers, and the number of zeroing around n image blocks by three convolution layers needs to be changed; The method comprises the following steps of: extracting motion change characteristic information of a target between frames in a plurality of time sequence ranges through a time-flow convolution neural network, and specifically, inputting m images into the time-flow convolution neural network; Then, an area of the target area in the previous frame which is concentrically and proportionally enlarged at least twice is taken, and an image is intercepted in the current frame; At last, the zeroing around m image blocks is completed by Numpy function, and finally the image features are output, wherein the time-flow convolution neural network includes eight convolution layers and five pooling layers, and it is also necessary to change the number of zeroing around m images by eight convolution layers;

[0029] The obtained image block feature information and motion change feature information weights are fused, a fully connected neural network is constructed according to the fused information, and the image blocks meeting the requirements are obtained, and then the frame regression operation is performed on the obtained image blocks to obtain the prediction information of the current frame target, wherein, the first fully connected layer and the second fully connected layer are cascaded in sequence, and B fully connected layer branches are deployed in parallel behind the second fully connected layer, and each fully connected layer branch serves as a third fully connected layer;

[0030] The fused feature information is used as the input of the fully connected neural network, and a two-dimensional vector is output through the calculation of the fully connected neural network. Among the image blocks of the spatial flow convolution neural network, the image block with the highest similarity score to the target is selected as the image block that meets the requirements.

[0031] The prediction information of the current frame and the determined prediction information of the previous frame are input into the corresponding trained correlation filtering convolution neural network, time flow convolution neural network and spatial flow convolution neural network models to obtain response mapping, and the Bagging algorithm is used to fuse all models to determine the final prediction information of the current frame.

[0032] According to the invention, through the convolution neural network of the correlation filter, the time-consuming offline training is not needed, and the time network and the space network are built on the basis of the correlation filter network, so that the time information and the space information of the target are further captured, and the accuracy of the algorithm is improved; At the same time, the convolution neural network is used to extract the motion change information of the tracking target between frames, which makes full use of the timeliness of video, reduces the influence of target occlusion and background noise, improves the accuracy of target position prediction information, and makes the tracking effect more accurate.

[0033] In the description of the present invention, it should be noted that the directions or positional relationships indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inside" and "outside" are based on the directions or positional relationships shown in the drawings, only for the convenience of describing the present invention and simplifying the description, and are not indicated or implied.

[0034] The above are only examples of the present invention, which do not limit the patent scope of the present invention. Any equivalent structure or equivalent process changes made by using the contents of the present specification and drawings, or directly or indirectly applied in other related technical fields, are equally included in the patent protection scope of the present invention.