Administrator's Guide for People Counting in Deep Learning NVR

Introduction

Overview

Synology Deep Learning NVR features powerful AI Image Analysis technology integrated into the Deep Video Analytics application within Surveillance Station. This functionality is provided without additional fees or external device installations.

This document guides users through the essentials of deploying People Counting cameras and suggests optimal installation scenarios for achieving high analytic accuracy.

The NVR's inbuilt GPU leverages Deep Learning AI for instant detection, tracking, and high-quality filtering of moving people. Upon detecting a mobile object, it calculates features of heads and shoulders to identify if the object is human.

Unlike traditional image analysis techniques that may produce false alarms, Synology People Counting utilizes deep learning for enhanced object recognition, anti-interference capabilities, and the classification of numerous object features, significantly improving human recognition accuracy and reducing miscalculations.

Synology Deep Learning NVR models support multi-channel image analysis, enabling tracking of multiple people simultaneously. For example, the DVA3219 can perform four People Counting analyses concurrently, tracking up to eight people per channel.

To ensure successful camera mounting and configuration, consider the following key factors:

• Design suitable installation scenarios.
• Select fitting camera models and locations.
• Assign appropriate software settings.

Mount Cameras

Select and Position Cameras

Camera model types and configurations significantly impact image quality. For People Counting, which relies on identifying human heads, clear and large images are crucial for accuracy.

General guidelines for camera setup include:

• For flexible stream quality, select ceiling mount cameras supporting multi-stream with 1920x1080@20 FPS or higher.
• Avoid panoramic or fisheye cameras, as image distortion can affect detection results.
• Maintain stable camera signals; wired connections are highly recommended.
• Keep camera lenses clean to prevent dust, insects, or stains from obstructing the view.
• Ensure passageways near the detection area are clear. Prevent visitors from lingering to avoid multiple counts. Remove mobile objects like automatic doors, escalators, and cleaning robots from the camera's field of view.

Camera Placement: Position cameras directly above entrances, facing straight down, ensuring complete human heads are captured in the footage. Arrange lenses so visitors pass through the camera's view horizontally or vertically.

Mounting Height: Mount cameras at least 2.5 to 4 meters (from floor to ceiling) above passageways. The height range and covered ground can be adjusted based on camera models and zoom settings.

The following table lists recommended focal lengths and corresponding heights for covering a 4-meter-wide passageway:

Prepare Suitable Lighting

Adequate lighting is critical. Insufficient light can cause blurry footage and lost details, while excessive illumination can lead to overexposure and reduced clarity.

Recommendations for lighting:

• Provide sufficient lighting, ideally over 300 lux, to ensure moving people are recognizable.
• Avoid direct sunlight in detection areas, as it can cause streaks or overexposure.
• Do not point lights directly at cameras to prevent overexposure.
• Camera night vision (IR) modes cannot compensate for insufficient ambient light; additional lighting may be necessary.
• Remove flickering or glowing objects, such as neon lights.
• Avoid uneven illumination, which can lead to missed detections in darker areas.
• Remove tilted light sources that cast shadows, as shadows can obscure human features.
• Adjust lighting color to match the environment, ensuring hair can be distinguished from clothing to prevent misidentification.

Configure Software Settings

After successfully mounting cameras, configure software settings in Deep Video Analytics (DVA) for optimal People Counting precision.

Begin with People Counting

People Counting functions by tracking head movements. The counter increments when the center of a person's head crosses the detection area.

Define the Detection Line

The detection line should be placed on the ground, centered in the camera view, and span the entire width of the passage. If the line is too short, people may pass without crossing it, leading to missed counts. The maximum line length is 4 meters.

Visual Description: A visual representation of a detection line drawn across a passage in the camera's field of view, indicating its position and coverage.

Select a Stream Profile

For optimal detection accuracy, select a stream resolution of at least 1920x1080@20FPS.

Visual Description: Software interface showing the 'Stream profile' setting, with options like 'Balanced (1280x720)' available for selection.

Edit the On-Screen Head Size

Accurate human head detection requires defining the on-screen head sizes. Within the 'Parameters' section, click 'Edit' and adjust the yellow object frame to set the appropriate head size.

Visual Description: Software interface displaying the 'Parameters' section, where a yellow object frame is adjusted to define the on-screen head size for accurate detection.

Improve Detection Accuracy

Beyond software settings, several factors can influence People Counting accuracy. This section outlines potential solutions, causes, and provides setup examples.

Select Proper Flooring

Simpler surroundings enhance People Counting's ability to analyze human features and generate accurate reports. Follow these flooring guidelines:

• If flooring has high reflection or sharp shadows, place a mat or carpet in the detection area.
• Use flooring that contrasts with the hair color of visitors; for example, light carpets for black hair and dark carpets for blonde hair.
• Opt for plain flooring to prevent complex patterns from interfering with analysis.

Note Possible Interferences

Despite careful planning, human heads may sometimes be undetected or misidentified. The following factors can cause miscalculations, though People Counting can still function:

• People under 120 cm may have heads too small to be identified, potentially filtered out by on-screen head size settings. Reducing the on-screen head size might increase interference from other small, moving objects.
• Outdoor camera accuracy can be affected by weather conditions like rain and snow, changes in shadows, or differences between day and night.
• People walking very close together may not be recognized correctly.
• People moving too quickly might not be detected.
• Individuals wearing hats, costumes, holding umbrellas, or accessories that cover their heads may be missed or affect the detection of others.
• Pets passing through the detection area can also affect calculations.

Setup Example

Key considerations for camera installation:

• Complete Heads: Human heads must be fully visible for high accuracy. If heads consistently appear at the screen edges, adjust camera height or use cameras with a wider angle of view.
• Frame Centering: Keep people passing through the middle of the camera frame for improved precision.

Visual Description: Illustrations demonstrating correct camera setup: one shows human heads appearing fully within the frame for high accuracy, while another shows heads at the edges of the screen, indicating a need for adjustment. A third illustration shows people passing through the middle of the camera frame for better precision.

For passageways wider than four meters, set up two cameras to ensure complete head capture and minimize size variations.

Visual Description: Visual guide for wide passageways, suggesting the use of two cameras to maintain complete head images and consistent sizing.

Collect Footfall Data

Once People Counting tasks are set up, you can begin collecting and tracking footfall data. This chapter details how to work with the People Counting feature.

Enable Crowd Detection

Crowd Detection sends event notifications and triggers alerts in Live View when the number of people in a location exceeds a predefined threshold. It is suitable for venues like stadiums and malls where footfall limits are necessary for safety. The headcount is dynamic, calculated by subtracting outgoing people from entering people.

Visual Description: Software interface for 'Crowd Detection' settings, showing an 'Enable' checkbox and a numerical threshold.

Generate Reports

After collecting footfall data, navigate to the 'Detection Results' page to generate a People Counting Report. Reports list entering and leaving people counts by date and time. The flexible display allows adjusting time units (e.g., hourly, daily) and viewing data from multiple tasks simultaneously. Reports can be exported as HTML files.

Visual Description: A bar chart displaying the number of entering and leaving people over time, with options to adjust the time unit and export the report.

Reset People Counter

The counter provides simultaneous two-way counting of people moving in and out of passageways. A schedule can be set to reset the counter automatically.

Visual Description: Software interface for 'Reset counter' settings, allowing users to schedule automatic counter resets on a weekly basis.

Visual Description: The 'Detection Results' page of the software, listing recorded people counting data with options to play, download, or delete individual entries.