Despite the rapid advancements in video analytics technology, accurate anomaly detection in video surveillance footage remains a significant challenge.
The use of video analytics for detecting anomalies is becoming increasingly popular across many industries, from security surveillance to manufacturing and beyond. With the ability to process vast amounts of video data in real-time and detect unusual events, video analytics holds the potential to significantly improve operational efficiency and security.
However, while video analytics has many benefits, it also poses significant challenges for accurate anomaly detection. In this article, we will explore the key challenges faced by video analytics in detecting anomalies and examine some of the strategies being developed to overcome these obstacles.
Defining anomaly
Despite the rapid advancements in video analytics technology, accurate anomaly detection in video surveillance footage remains a significant challenge. As Fabiola Ruvalcaba, Commercial Lead for Video Analytics at Genetec points out, the difficulty lies in defining what constitutes an abnormal event.
While detecting outliers in a two-dimensional dataset can be relatively straightforward using unsupervised learning techniques, the complexity, and variability of video surveillance footage pose unique challenges. In this context, finding relevant anomalies requires sophisticated algorithms and approaches that can distinguish between genuine threats and false positives.
“Anomaly detection in video surveillance has been tried many times, and there is still no viable solution,” said Ruvalcaba. “The problem is the definition of what is abnormal. If the problem is very constrained, such as in a two-dimensional dataset like crime statistics in an area, it is easy to detect abnormal data using unsupervised learning. However, events in video surveillance footage can vary greatly. So, it is hard to find relevant outliers.”
Quality of training data
The effectiveness of a deep learning system for anomaly detection is directly linked to the quality of its training data. Developing accurate and comprehensive training sets can be a time-consuming and resource-intensive process. Moreover, ensuring that the system can accurately identify anomalies such as smoke or liquid spills under varying lighting conditions presents a significant challenge.
To achieve this level of precision, extensive data sets that account for factors such as movement, and the physical properties of materials must be used. In other words, relying solely on learned images may not be sufficient, and a more holistic approach that considers multiple factors is necessary.
“These are the kinds of advances we can look forward to as the neural networks evolve and improve over time,” said Rui Barbosa, Product Manager, i-PRO Americas. “i-PRO has recently introduced Scene Change detection, which allows integrators and even end users to teach a camera what a normal scene looks like. Users can capture multiple images over time to represent the scene in different lighting environments.”
Areas can then be specified to watch for any anomalies in the scene from normal. This is an example of a completely new and powerful use of AI that can do everything from monitoring a no-parking zone to alerting staff when stock on a shelf drops below a preset threshold.
John Rezzonico, CEO of Edge 360, also added that data quality is crucial, as the accuracy of the analysis relies on the quality of the data captured by the video cameras. Therefore, ensuring that high-quality data is captured is essential to avoid inaccurate results and false positives.
Lighting conditions
Rezzonico pointed out that another significant challenge is lighting conditions. Variations in lighting, such as changes in intensity, shadows, and reflections, can distort the images captured by cameras, making it difficult for algorithms to identify and track objects accurately. These challenges may be particularly acute in outdoor environments, where lighting conditions can be highly variable and unpredictable.
“Therefore, lighting must be optimized to ensure the best possible results. Camera placement is also crucial as cameras need to be strategically placed to capture critical areas and minimize blind spots,” Rezzonico said. “Network bandwidth is also vital, as high-quality video requires large amounts of bandwidth, and a lack of bandwidth can result in video lag, poor image quality, and inaccurate analysis.”
Need for advanced algorithms
According to Greg Skarvelis, Director of Solutions Design and Chief Solutions Architect at Intellicene, some of the most significant challenges include the complexity of video data, variations in lighting and environmental conditions, object detection and tracking, real-time latency, and privacy concerns.
These challenges can make it difficult to identify and track anomalous behavior effectively, highlighting the need for advanced algorithms and approaches that can account for these factors. By developing sophisticated solutions that can overcome these challenges, organizations can unlock the full potential of video analytics and improve their operations in a range of applications.
“Overcoming these challenges requires the use of advanced algorithms, robust sensors and cameras, powerful computing resources, and specialized hardware and software solutions,” Skarvelis said. “Addressing these challenges can significantly improve the accuracy of analytics, making it an essential tool for a wide range of applications.”
Finding the ideal VCA solution
According to Victor Hagelbäck, CMO/CPO of Irisity, it all comes down to the quality of AI and video analytics platforms. With the right system in place, you can accurately detect anomalies even in the most difficult environments.
“The best anomaly detection systems in the market are based on accurately classifying every object and every motion over long periods of time,” Hagelbäck said. “Anomalies can then be detected by comparing what is happening right now to what is normal in the view of that specific camera.”
The most advanced systems are designed to accurately classify objects and motions over extended periods, allowing for more precise anomaly detection. By comparing the current situation to what is normal within the view of a specific camera, these systems can quickly identify and alert operators to potential threats. With the right platform in place, organizations can better protect their assets, reduce risk, and improve overall operational efficiency.