The invention discloses a computer-vision-based security detection method, which is used for meeting the requirements of current security detection on persistence and high accuracy. The method comprises the following steps: performing original image gray processing on original video data, performing color space standardization on an input image, and regulating the contrast of the image to reduce the influence of a local shadow of the image and an illumination change and suppress the interference of noise; describing characteristics of a pedestrian by virtue of a gradient histogram, classifying by combining an SVM (support vector machine), analyzing a behavior of the pedestrian by virtue of deep learning after the pedestrian is found out, and sending security alarming information if the behavior of the pedestrian is inconsistent with a security specification. According to the method, existing hardware equipment can be fully utilized, so that changes in an original system are maximally reduced. Moreover, more image details can be understood by deep learning, so that a higher recognition rate is achieved. A deep learning neural network is less sensitive to the environment, light and noise, and can run in each environment after being trained once, so that high generalization capability is achieved.