1. Current status of video surveillance systems and its AI technology

As our cities grow and public spaces become busier, ensuring the safety of people in places like shopping centers, parks, and hospitals has never been more critical. Public facility operators are responsible for preventing incidents caused by abnormal behavior, which can lead to accidents or even serious safety threats. Traditional video surveillance systems are often inadequate, as they mainly serve to document events after they occur rather than actively preventing them.

In these scenarios, we face several pressing issues:

Surveillance systems do not actively monitor in real-time and cannot alert to abnormal behavior.
After abnormal behavior occurs, it only records and observes for analysis, capturing and storing video images, revealing what happened after the fact.
A large amount of human and financial resources is required.

Now, suppose we propose such an intelligent abnormal behavior detection and monitoring system:

Continuous 24-hour real-time detection, where the camera automatically analyzes human behavior and identifies effective patterns from captured image data.
When abnormal behavior occurs, the system can alert security personnel to take timely action, thereby preventing crimes and other incidents.

Then the model of the practical application of this technology abnormal behavior we categorized as follows:

Scenario 1: In a home setting, an elderly person falls accidentally when their family member are not around. The home surveillance camera can immediately alert their family.

Scenario 2: On a street, a fight breaks out and a number of onlookers. The road surveillance camera can promptly alert and notify the authorities.

Scenario 3: In a hospital, a patient remains seated for a long time, with eyes closed and unmoving. The hospital’s surveillance camera can quickly alert and notify medical staff.

…and so on.

2. Challenges and Solutions in Abnormal Behavior Detection Technology

Although some classic algorithms have achieved good results in abnormal behavior detection, but still face the following difficulties:

The high complexity and variability of crowd behavior make abnormal behaviors diverse, requiring high recognition capabilities from behavior models.
How to ensure the accuracy of detecting abnormal behaviors of crowds in complex scenes such as occlusion and light changes;
Many algorithms are computationally intensive and have low accuracy, and achieving a balance between real-time processing and detection accuracy is a research challenge.

Maadaa.ai proposes the following solutions in the field of abnormal behavior detection:

1. Key Frame Selection

Collect data under various lighting conditions, at different times, and from different distances, involving people of different age groups. Collect positive and negative samples in a 1:1 ratio under normal scenarios, while also defining specific abnormal behaviors.

2. Key Frame Selection

Use training datasets containing only normal behaviors to establish normal behavior models. Extract deep features from continuous video frames to fully represent the appearance and motion information of crowd behavior in the video.

3. Cross-Scenario Abnormal Behavior Tracking and Detection

Existing methods typically train normal behavior models for a single fixed scenario, making it difficult to detect abnormal behaviors across different scenarios. In our data collection process, we capture abnormal behaviors from multiple cameras and angles, annotating key frames with the same ID.

4. Generalization and Detection Capability of Normal Behavior Models

Abnormal behavior detection in crowds requires the ability to identify rare yet seemingly normal behaviors. These behaviors are rarely present in conventional training video datasets but are often the source of false alarms. An ideal normal behavior model should cover all regular behaviors in the training videos and identify any deviations from this model.

5. Definition and Description of Abnormal Behavior

The scope of abnormal behavior is broad and lacks a precise definition, making it difficult to understand its underlying causes. By combining abnormal behavior with its environment and context information, we can generate concise textual explanations of abnormal behavior, which will help us determine whether the abnormal behavior is truly a false alarm.

3. Maadaa.ai’s Case Study

— Customer Abnormal Behavior Monitoring Data Collection in a Large Shopping Mall

Application Scenario: Customer safety monitoring in the escalator area of a large shopping mall.

Current escalators in shopping malls have safety issues, making it difficult to detect unsafe behavior in real-time and take preventive measures, leading to significant safety hazards. Furthermore, the management efficiency of escalator operators is low; in emergencies, they cannot alert timely.

AI Solution

The escalator AI behavior recognition system achieves full coverage, with no blind spots, of the entire escalator area through real-time video monitoring. It captures the real-time status of the escalator, improving property management efficiency. The AI system accurately identifies uncivilized and unsafe behaviors on the escalator, automatically alerts in emergencies, and notifies nearby rescue personnel for quick response.

Challenges in Data Collection

Difficult to capture data of diverse escalator types.
Rich shopping mall environments are required.
Using a model for fall detection is risky.
It is difficult to define the behavioral actions of real falls at different locations of escalators of different heights.

Maadaa.ai’s Customized Data Collection Solution

Diverse Data Collection Scenarios to Ensure Data Variety

Various types of escalators and installation environments.
Sampling and simulation scheme design for real abnormal behavioral data (escalator fall, etc.)

2. Establish Data Collection Standards to Ensure Consistency

Refine data collection scenarios, such as escalator status (moving/stationary), event location (escalator entrance/middle-lower section), abnormal behavior (backward fall/forward fall), and carried objects (luggage/backpack/bag).
Refine the data collection plan, including video capture volume, actor action execution standards, and carried objects.
Deploy specialized equipment, including the position, direction, and height of multiple cameras.