Anomaly Detection Solution for a Gaming Rooms Owner

Client

A US-based owner of gaming rooms located throughout the country

Industry

Entertainment

Services

Product development, AI, ML

Tech

Java, Python

Challenge

The customer, a US-based owner of 20+ amusement arcades located throughout the country, wanted to enhance safety in their gaming rooms, reduce slot machine downtime, and optimize maintenance. For that, they sought to craft an anomaly detection solution that would integrate into their current video surveillance system and analyze video footage in real time to detect inappropriate gamer behavior, such as physical violence or slot machine damage, detect abandoned items, and timely report slot machines that went out of order. Having reviewed the proposal devised by the ITRex anomaly detection team, the customer entrusted us with crafting the AI-driven anomaly detection solution and integrating it into their cloud-based monitoring system.

In particular, we were tasked with:

●

Choosing an optimal algorithmic solution to the customer’s anomaly detection challenge

●

Creating a training data set and training anomaly detection algorithms at the base of the solution

●

Integrating the trained and tuned anomaly detection solution into the customer’s cloud infrastructure

Solution

●

The anomaly detection solution accurately detects sequential events, such as physical violence or damage inflicted on slot machines, and slot machines that went out of order, timely informing the monitoring center of the detected anomalies.

●

The solution uses variational autoencoders (VAE) to detect sequential anomalies, like fights or inflicting damage. To train high-performing VAEs, the ITRex team generated a dataset of 150 videos featuring said anomalies and trained the VAE on this data.

●

The OpenCV framework was used for preprocessing the videos (think: reading frames and resizing them). The ITRex team also employed the torchvision library for data normalization and augmentation. We developed the network architecture and carried out the training process relying on PyTorch. To accelerate the model's inference speed, we converted the model to the TensorRT format.

●

During training, the VAE learned to reconstruct the input videos. The closer the reconstruction to the input, the lower the error rate. The VAE is then applied to new, unseen data. If it shows a low error rate, we conclude that footage features an anomaly.

●

The solution relies on cross-validation to detect machines that are not functioning properly. To do so, the solution detects a system error message that appears on a machine’s screen and analyzes it against an error template based on feature-based matching.

●

The anomaly detection solution seamlessly integrates into the customer’s cloud-based monitoring system. Whenever an anomaly is detected, the solution generates alerts for the customer’s personnel to take an appropriate action.