41 results about "Motion History Images" patented technology

Methods of processing video are presented to generate signatures for motion segmented regions over two or more frames. Two frames are differenced using an adaptive threshold to generate a two-frame difference image. The adaptive threshold is based on a motion histogram analysis which may vary according to motion history data. Also, a count of pixels is determined in image regions of the motion adapted two-frame difference image which identifies when the count is not within a threshold range to modify the motion adaptive threshold. A motion history image is created from the two-frame difference image. The motion history image is segmented to generate one or more motion segmented regions and a descriptor and a signature are generated for a selected motion segmented region.

The invention provides an image-type fire detection method based on a cumulative prospect image, builds the cumulative prospect image, and is used for real-time detection of a fire image. The method comprises the following steps of: extracting the cumulative prospect image obtained by a monitoring camera through a computer; blocking the image; counting brightness values of all pixels of each image block in the cumulative prospect image; judging according to preset sensitivity; and sending out a warning signal if a fire is detected, otherwise returning to the initial steps to continuously process next frame image. According to the method, the definition of historical motion images is changed to obtain the cumulative prospect image, and the cumulative prospect image can be used for reflecting active characteristics of fire flame better and greatly reducing false alarm rate and has good anti-noise capability. Moreover, the method has the advantages of simpler arithmetic principle, low computation and quite good real-time performance and can meet the requirement of the conventional image-type fire detection technology for the real-time performance.

A gesture judgment method used in an electronic device having frame capturing function is provided. A plurality of MHI (motion history image) angles / directions are obtained from a plurality of corresponding MHIs. Whether a current gesture control is valid is judged according to the MHI angles / directions. If the current gesture control is valid, then weight assignment is performed on the MHI angles / directions to obtain a judgment result of the current gesture control.

The present invention relates to a flame detection method and system, wherein the method includes: using camera monitoring equipment to capture video stream images of fire monitoring areas, and extracting motion history images from each frame of video stream images; Carry out class flame color pixel extraction, obtain class flame color image; Carry out logical AND operation to described motion history image and class flame color image, screen out the flame region candidate block; Screen the flame region candidate block of several frame continuous video stream images, And obtain the flame beating times and the edge beating amplitude value of the flame area candidate block of each frame video image; According to the flame beating number of times and the edge beating amplitude value, calculate the probability that each flame area candidate block is a flame area; when the probability If the preset value is exceeded, it is determined that a fire occurs in the fire monitoring area corresponding to the flame area candidate block, which improves the accuracy of fire detection.

The invention discloses a real-time intelligent monitoring method based on natural video. The method uses the knowledge of computer image processing and artificial intelligence and realizes unmanned intelligent monitoring and alarm to the action of pedestrian in public places and important sensitive places. Firstly, video frame serial sections which need to be studied are extracted, movable historical images are obtained, which reflect the movement process of the people; on this base, the user-defined method for extracting the eigenvector is used, vector representation of the specific movement series is obtained, and the vector sample is stored in the sample database; as for the video frame serial which need to be monitored, the eigenvector and sample data are mapped in the low dimensional space, the corresponding classification by the optimized method is obtained and alarm is carried out. Owing to the sample study mechanism and the classification mechanism of the designed actions in the text, the method of the invention improves the accuracy of identification and strengthens the expansibility of identification; by designing the eigenvector representation and extraction method of movement serial of the people, the completeness and accuracy of action representation are strengthened.

The invention discloses a view-independent human action identification method based on template matching, which can identify a plurality of pre-defined typical actions in a video. When constructing a template, a motion history image under a plurality of projection viewpoints are calculated for each sample action and polar coordinate characteristics are extracted, the polar coordinate characteristics are mapped to a low-dimensional sub-space by adopting a manifold learning method, and super balls are constituted for the sample actions in the sub-space on the basis of the low-dimensional coordinate of the multi-viewpoint polar coordinate characteristics. An action template is composed of a plurality of super balls with known ball centers and radiuses. When an unidentified action is given, the motion history image and the corresponding polar coordinate characteristics of the action are firstly calculated, then the polar coordinate characteristics are projected into the template action sub-space to obtain the low-dimensional coordinate, the distances from the coordinate to all the ball surfaces of the super balls are calculated, and the nearest super ball is selected as the identification result. The technology provided by the invention realizes the view-independent action identification and has higher application value in the video monitoring field.

The invention discloses a fall behavior recognition method based on a three-dimensional convolutional neural network. Firstly, a fall data set video is obtained and preprocessed to obtain a fall behavior video sample; the video is combined with a three-frame difference method based on a mixed Gaussian and an adaptive threshold. The target detection method removes the background, and then uses small area removal and morphological methods to obtain the complete human target area; extracts the optical flow movement history image features of the human target area, and then increases the sample set by means of data overlapping and amplification for the feature image; The falling behavior sample set after overlapping and amplifying is randomly divided into a training sample set and a verification sample set according to a ratio of 7:3 to input a 3D convolutional neural network model classifier and continuously iteratively train, while using the verification sample set to continuously verify the model classifier; The test sample set is input into the trained model classifier to complete the fall behavior recognition. The invention solves the problem of low classification recognition rate and precision caused by background interference in the existing fall detection method.

The invention discloses a behavior recognition method based on a motion history image and a convolutional neural network, comprising the following steps: S1, acquiring an input original video image, and processing it through a behavior sequence feature extraction method based on a motion history image; S2 1. Using a method based on a deep convolutional neural network to conduct behavior recognition on local motion history images to obtain a behavior category classifier, and finally output the behavior recognition result through the behavior category classifier. The present invention calculates motion history images from original video sequences, which not only reduces the amount of information to be processed, but also extracts key spatio-temporal information in behavior recognition; then uses motion history images as input to build a deep convolutional neural network, and then uses Stochastic gradient descent (SGD) and Dropout strategy train the network, and finally realize the classification of behavior categories. The invention can be effectively applied to online real-time behavior recognition.