Video Anomaly Detection Using Patch Similarity Neural Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current video anomaly detection methods struggle to distinguish rare but normal activities from abnormal activities in surveillance videos, leading to false positives and missed anomalies due to computational inefficiencies and the difficulty in collecting diverse abnormal activity data.
Innovation Solution
A neural network-based system that compares video patches from input videos directly with those from a database of normal activities, trained to classify similarities and dissimilarities, allowing for efficient detection of anomalies without requiring examples of abnormal activity, and optimizing memory usage by ensuring all training patches are dissimilar.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If aggregation-based anomaly detection methods are used to model normal motion patterns, then anomaly detection capability is improved, but computational cost increases and rare normal motions are missed
Solution Approach 1:
The video is segmented into small patches that are compared directly against training patches. This segmentation approach avoids the need to compute complex motion aggregations across entire video sequences, thereby reducing computational cost while maintaining anomaly detection capability through direct patch-level comparison.
Solution Approach 2:
Instead of modeling what normal motion looks like and detecting deviations from that model, the invention inverts the approach by directly comparing test patches against a database of training patches. Anomalies are detected when no sufficient match is found, eliminating the need for complex motion aggregation and probability distribution modeling.
2Device complexity
If aggregation-based methods are used to represent normal motion, then anomaly detection is simplified, but rare normal activities are misclassified as anomalies
Solution Approach 1:
A comprehensive database of training video patches is prepared in advance, covering diverse normal activities including rare ones. During detection, test patches are directly compared against this pre-prepared database, ensuring that rare normal activities are recognized as such rather than being misclassified as anomalies.
Solution Approach 2:
The invention uses direct copying and comparison of video patches between training and test sequences. Instead of creating abstract motion aggregations that lose detail, the actual video patches are stored and compared directly, preserving fine-grained details of rare normal activities and enabling accurate distinction between normal and anomalous behaviors.
3Productivity
If direct comparison methods are used between input and training video, then computational efficiency is improved, but the ability to handle diverse normal variations is reduced
Solution Approach 1:
The invention extends direct comparison from simple spatial patch matching to spatio-temporal patch matching by incorporating temporal dimensions. Video patches include multiple frames that capture motion patterns, allowing direct comparison to handle diverse normal variations while maintaining computational efficiency through the efficiency of direct similarity measurement in the spatio-temporal domain.
Data Source
AI summary
A system for video anomaly detection includes an input interface configured to accept an input video of a scene, and a memory configured to store training video patches of a training video of the scene capturing normal activity in the scene, and store a neural network trained to compare two video patches to declare the compared video patches as similar or dissimilar. The system also includes a processor configured to partition the input video into input video patches, compare, using the neural network, each input video patch with corresponding training video patches retrieved from the memory to determine if each input video is similar to at least one corresponding training video patch, and declare an anomaly when at least one input video patch is dissimilar to all corresponding training video patches.