Research On The Algorithm Of Escalator Pedestrian Anomaly Detection Based On Image

Aiming at the frequent accidents and dangerous behaviors on the escalator,the image processing was used to detect and to warn the abnormal behaviors,such as carrying pets,trolleys,holding children and climbing on the escalator in real time.On this basis,an intelligent monitoring system was developed to detect and to analyze the escalator scene in real time to give an online alarm to protect the safety of passengers.The following aspects was mainly studied.Firstly,based on the problem of target motion blur,caused by the rapid movement at the escalator entrance and exit,the generated countermeasure network was used to reconstruct the blurred image to reduce the motion blur.The feature was extracted by 3 * 3 convolution in generator network model.The dense residual blocks were introduced to enrich the global and local information between feature graphs.During the up-sampling,bilinear interpolation was used to avoid the uneven distribution of pixels.In order to improve the definition of deblurring image edge,L1 norm of gradient image difference between blurred image and real image,was used as penalty term in generator loss function.The Tiny YOLOv3 was trained and learned to detect the three kinds of abnormal objects,such as taking pets,baby strollers and Pedestrian falls.The deblurring improved the detection accuracy from 66.21% to 74.37%,and the recall rate from 42.30% to 51.85%.Secondly,it is difficult to extract foreground when the target is crowded and the color is close to the background.Escalator Pedestrian Abnormal Behavior Detection Method Based on SVM was proposed.HOG+SVM classifier was selected to extract features for three kinds of abnormal objects,such as pedestrian falling on escalator,baby stroller and carrying pet.Two kinds of SVM classifiers were selected in turn to train.The category with the most votes by voting was taken as the final detection result.The experimental results showed that the algorithm of HOG+SVM classifier could detect abnormal objects on escalators.The detection accuracy was 81.71%.The missed detection rate was 31.68%.And the average detection time of a frame was 800 ms.It was difficult to meet the detection requirements of the actual escalator scene.Neural network was combined with the help of feature fusion to improve the performance of the HOG+SVM classifier.The deep learning was selected to do the further research.Thirdly,because the traditional algorithm is difficult to meet the requirements of real-time detection of escalator pedestrian abnormal behavior.Escalator pedestrian abnormal behavior detection algorithm based on improved Tiny YOLOv3 is proposed.The K-means++algorithm,together with lower randomness,was selected to optimize the parameters of the prior box.The 18 layers of depth separable feature extraction,was used to minimize the calculation when the network was deepened.Pyramid feature was used to hold the scale unchanged,caused by different distance between abnormal target and camera.The scale prediction was also improved.The three SE-Resent module was added to the extracted scale features.Thus,the effective features was strengthened and the invalid features was weakened.Soft-NMS strategy was selected to reduce the target of the missed detection,caused by NMS false filtering during the in the test.Finally,GPU was used for multi-scale training to get the optimal weight model.Compared with Tiny YOLOv3,the optimized model reduced the average miss detection rate by 23.1%,improved the detection accuracy by 4.0%.It gives the better consideration to the detection accuracy and to the real-time performance.Fourthly,based on the above research results,the escalator abnormal behavior intelligent monitoring system was designed and implemented.The actual test results showed that the escalator abnormal behavior detection system had good robustness.The detection accuracy and real-time performance could meet the actual needs.