Camera Anomaly Detection Using Normal-Data Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing surveillance systems face inefficiencies in anomaly detection due to sensitivity to lighting changes, camera geometry, and lack of training data for rare events, leading to false alarms and reliance on human oversight.
Innovation Solution
A sequence of machine-learning models, including a pre-processing model to decorrelate input data and an autoencoder, to enhance robustness and accuracy in anomaly detection, allowing for automatic calibration and reduced false positives/negatives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional supervised multi-category learning with deep networks is used for anomaly detection, then the system can process video data, but it suffers from shortage of training data for anomalous events and produces many false alarms
Solution Approach 1:
Instead of training the network to detect anomalies directly (supervised learning requiring anomaly labels), the patent inverts the approach by training the network to reconstruct normal video frames using autoencoders. Anomalies are detected when the reconstruction error exceeds a threshold, eliminating the need for annotated anomaly training data while maintaining detection capability.
Solution Approach 2:
The patent introduces an autoencoder as an intermediary component between the input video frames and the anomaly detection decision. The autoencoder learns to compress and reconstruct normal patterns, serving as a mediator that implicitly defines what is 'normal' without requiring explicit anomaly examples, thereby solving the training data shortage problem.
2Quantity of substance
If autoencoder-based approaches are used for anomaly detection, then training can be performed without anomaly footage, but the models are not robust against lighting changes, geometry changes, or camera movements
Solution Approach 1:
The patent applies data augmentation as a preliminary action during the training phase. By artificially transforming normal video frames with various lighting conditions, geometric transformations, and camera movements before feeding them to the autoencoder, the model learns to be invariant to these changes, thereby improving robustness without requiring additional real-world training data.
Solution Approach 2:
The patent employs data augmentation techniques that systematically change parameters such as brightness, contrast, rotation angles, and translation distances during training. This exposes the autoencoder to a wide range of parameter variations, enabling it to maintain performance across different environmental conditions and camera configurations.
3Extent of automation
If deep network approaches are used for video anomaly detection, then automated detection is achieved, but the systems require continuous lighting and do not work well in the presence of windows or changing weather
Solution Approach 1:
The patent applies lighting-related data augmentation (brightness adjustment, contrast variation, shadow simulation) as a preliminary action during training. This prepares the autoencoder to handle diverse lighting scenarios including nighttime, backlighting from windows, and weather-related lighting changes, enabling automated detection to function reliably across varying environmental conditions.
4Ease of manufacture
If traditional computer vision algorithms are used for anomaly detection, then the system can run with existing feeds and provide basic functionality, but pixel-level changes trigger alarms leading to false positives and user system disablement
Solution Approach 1:
The patent replaces traditional mechanical computer vision algorithms (pixel comparison, edge detection, motion detection) with a data-driven deep learning approach using autoencoders. This substitution enables the system to learn complex normal patterns and distinguish them from anomalies more effectively, significantly reducing false alarm rates while maintaining implementation feasibility through available frameworks.
Data Source
AI summary
Examples relate to an apparatus, a method and a computer program for detecting an anomaly in input data, to a camera device and a system comprising such an apparatus, and to a method and computer program for training a sequence of machine-learning models for use in anomaly detection. The apparatus for detecting an anomaly in input data is configured to process the input data using a sequence of machine-learning models. The sequence of machine-learning models comprising a first machine-learning model configured to pre-process the input data to provide pre-processed input data and a second machine-learning model configured to process the pre-processed input data to provide output data. The first machine-learning model is trained to transform the input data such, that the pre-processed input data comprises a plurality of sub-components being statistically independent with a known probability distribution. The second machine-learning model is an auto-encoder. The apparatus is configured to determine a presence of an anomaly within the input data based on the output of the second machine-learning model.