Anomaly Detection In Crowd Scene Surveillance Video Based On Transfer Learning

Abnormal event detection in video surveillance is extremely important,especially for crowded scenes.In recent years,many algorithms have been proposed based on hand-crafted features.However,it still remains challenging to decide which kind of feature is suitable for a specific situation.In addition,it is hard and time-consuming to design an effective descriptor.In this paper,video's spatial-temporal feature is extracted by a 3D convolutional neural network model and we use isolation forest algorithm for anomaly detection.Firstly,the algorithm to extract video feature is based on deep learning method.We use a pre-trained network structure to extract the spatial-temporal feature.After going through the feature extract model,every overlapping perceptual field would produce a feature vector.This vector will represent the spatial-temporal information extracted from the perceptual field.According to our experience,3D convolution network can give us a better result than 2D CNN.Secondly,we use the isolation forest algorithm to detect the abnormal feature vectors which represent the perceptual field.Isolation forest is a machine learning algorithm based on ensemble learning.It has high accuracy and low complexity so that it can satisfy the requirement of computation for large data.We use it for the task of crowd anomaly detection and prove that it is a suitable algorithm for this task.Thirdly,we propose an algorithm to localize the abnormal event in the video.We use a Gaussian function to process the probability of the feature map to get the probability map of the perceptual field.After that,we add the probability map of perceptual fields to get the probability map of the original video.As a result.this paper propose novel method for video feature extraction,anomaly detection and anomaly localization.Experimental results show that the deep model is effective for abnormal event detection in video surveillance.