Abnormal Detection Of Station Based On Deep Learning

A smart station is an intelligent transportation hub,and station anomaly detection is a key technology in smart stations.To achieve the detection of smoking,falling,and open flames in stations,this proposes an algorithm SFF-YOLO(Smoke-Fall-Fire-YOLO)based on deep learning.The model is deployed on Jetson Nano,and the model inference is optimized.Finally,the detection is carried out through the station anomaly detection system.Specific work and research are as follows:(1)The special data sets for the detection of smoking,falling and open flame are constructed.To address the limited availability of station anomaly detection datasets,a dedicated Smoke Fall Fire Dataset is constructed by searching online and collecting videos shot from different distances and angles.Frames are extracted from these videos,sorted and labeled to construct the dataset.The dataset contains a total of 2400 images with a resolution of640×640 pixels,including 900 images for smoking and falling respectively,and 600 images for open flames.(2)A new algorithm SFF-YOLO is proposed to meet the requirements of accuracy and real-time performance for station anomaly detection of smoking,falling,and open flames.Based on YOLOv7-tiny,the Sim AM attention mechanism is embedded SFF-YOLO in the backbone layer,and SPPFCSPC(Spatial Pyramid Pooling Fast Cross Stage Partial Construction)is used at the connection between the backbone and neck layers to reduce the computational and parameter overhead caused by the parallel Max pooling structure in the original SPPCSPC model.In addition,GCT(Gated Channel Transformation)is introduced into the neck layer to gain the deep convolutional network global context information.The comparative experimental results of the SFF-YOLO algorithm on the Smoke Fall Fire Dataset demonstrate that the detection accuracy of smoking,falling,and open flames by the SFF-YOLO algorithm is 95.4%,99.6%,and 83.9%,respectively.This represents an improvement of 6.7%,0.2%,and 6.6% over YOLOv7-tiny,with a reduction in parameters of 1.08%.The FPS is 143 frame /s.(3)A station anomaly detection system based on Jetson Nano has been implemented.It uses Python language and PyQt5 framework to design the system front-end interface and implement the back-end logic functions,as well as encapsulate the algorithm.After configuring the software environment of the Jetson Nano embedded platform,the model is deployed and optimized for inference.The comparison experiment results of the SFF-YOLO model before and after TensorRT inference optimization on Smoke Fall Fire Dataset showed that the FPS increased from 9.8 frame /s to 16.7 frame /s under the FP16 accuracy without reducing the detection accuracy.With frame skipping detection,real-time anomaly detection at a train station can be achieved.