High-altitude falling object detection method based on visual Transform

Abstract

The invention relates to a high-altitude falling object detection method based on a visual Transform. The method comprises the following steps: S1, obtaining a video image of a monitoring area; S2, extracting features of a current frame by using a convolutional neural network according to the acquired video image, and splicing the features with features extracted from a previous frame; S3, inputting the spliced features into an encoder, and inputting a result into a decoder after calculation; S4, embedding the characteristic value of the historical frame into an input decoder, and performing correlation calculation on the output of the encoder and the characteristic value of the historical frame; S5, respectively inputting the last sequence obtained by operation into the three full-connection networks, and further calculating the category probability that the object is a falling object, a target bounding box and an inter-frame motion vector of the object; and S6, identifying a falling object according to a calculation result, tracking the falling object and sending out an alarm prompt. According to the invention, the transformer technology is applied to falling object detection, and the problems of hysteresis and high false detection rate in a traditional multi-stage falling object detection method can be solved.

Description

Technical field [0001] The present invention relates to image recognition techniques, in particular a high-altitude pendant detection method based on visual transformer. Background technique [0002] With the more built in the high-rise building, the harm of high-altitude pendants is also increasing, and the high-altitude pendant incidents that occur in the news report are also increasing trend. Due to the high hazards of high-altitude treasures, serious threats of road pedestrians' life safety Therefore, it is important to detect the timely detection of high-altitude pendants. At present, conventional detection methods are typically detecting objects in the picture, and then tracking the same object in the multimrace image, and finally determined whether or not a parabolium is judged by a trajectory fitted in each frame in each frame. The method is integrated. Objective detection, target tracking, and trajectory judgment. Among them, the target detection algorithm can probably b...

AI Technical Summary

Problems solved by technology

Both of the above two object detection algorithms have certain defects. Among them, the robustness of the traditional image processing algorithm is low, and it is easily interfered by external factors such as light, noise, clarity, etc., resulting in large missed and false detections. And it usually consumes computing resources, and it is difficult to meet the real-time requirements; while using the deep detection network for target detection can greatly improve the detection accuracy, but in the tracking and trajectory judgment stage, the step-by-step high-altitude parabolic detection algorithm Only after obtaining the complete trajectory of the object can it be judged whether it is a falling object, which makes the detection have a certain lag. In addition, the error of object detection and tracking will be directly superimposed on the trajectory judgment, which greatly affects the accuracy of the algorithm.

Images