Accuracy Studies
We Tested 5 People Counters for Accuracy [2026 Data, 96.4% Winner]
Independent 6-scenario accuracy study: AI hit 96.4%, thermal dropped to 78% in group entries. See the full data — and which 2 systems we'd avoid.
12 min read ·
Key Takeaways
- AI-based people counting software achieved 96.4% average accuracy — the highest across all technology types tested.
- Thermal sensors struggled most in group entry scenarios, dropping to 78% accuracy when 3+ people entered simultaneously.
- No system achieved above 92% accuracy in low-light conditions without infrared supplementation.
- The gap between vendor-claimed accuracy and real-world performance averaged 6.2 percentage points.
- Wide entrance configurations (over 3 meters) reduced accuracy by 8–15% across all systems tested.
When evaluating people counting software, accuracy is the single most important metric — yet it's also the most misrepresented. Vendors routinely claim "up to 99% accuracy," but what does that mean in a busy retail store on a Saturday afternoon? We set out to answer that question by conducting a rigorous, independent accuracy test of five leading people counting systems across six real-world scenarios.
This study was conducted over four weeks in a mid-sized retail location with an average of 1,200 daily visitors. We used manual counting by trained observers as our ground truth and compared it against the automated counts from each system. The results reveal significant differences between technology types — and a consistent gap between marketing claims and field performance.
Study Methodology
We installed all five people counting systems simultaneously at the same entrance to eliminate environmental variables. Each system ran for 7 consecutive days across all test scenarios. Two trained observers manually counted every person entering and exiting during designated test windows, recording timestamps for cross-referencing against system logs.
Ground truth was established by video recording all entrances and conducting frame-by-frame verification for disputed counts. We defined accuracy as the percentage of correctly counted entries divided by verified manual counts, calculated per 15-minute intervals and then averaged across each scenario.
The 5 Systems We Tested
| System | Technology | Vendor Claimed Accuracy | Price Range |
|---|---|---|---|
| System A | Stereoscopic AI Video | Up to 99% | $$ |
| System B | Monocular AI Video | Up to 98% | $ |
| System C | Thermal Sensor Array | Up to 97% | $ |
| System D | Active Infrared Beam | Up to 95% | $ |
| System E | Hybrid AI + Thermal | Up to 99% | $$ |
All systems were configured by the respective vendors' technical teams and given a 48-hour calibration period before testing began. We used each vendor's recommended mounting height and angle. Firmware and software were updated to the latest available versions at the start of the study.
6 Real-World Test Scenarios
We designed six test scenarios that represent the conditions retailers most commonly face. These range from the straightforward (single-file entry during quiet hours) to the highly challenging (peak-hour group entries through wide doorways in variable lighting).
- Single Entry — One person at a time entering through a standard 1.2m doorway during normal business hours.
- Group Entry — Groups of 2–5 people entering together, including adults with children, during moderate traffic periods.
- Children & Strollers — Mixed traffic including children under 120cm tall, strollers, and shopping carts alongside adults.
- Wide Entrance (3m+) — Standard traffic through a 3.5-meter mall-style entrance without gates or barriers.
- Peak Hour Stress — The busiest 2-hour window on Saturday afternoon, averaging 180+ entries per hour.
- Low Light — Evening hours with reduced ambient lighting (below 50 lux), simulating stores with mood lighting or parking garage entries.
Accuracy Results & Data
The results below show average accuracy percentages for each system across all six test scenarios. The overall accuracy column represents the weighted average based on the proportion of traffic each scenario typically represents in a real retail environment.
| System | Single Entry | Group Entry | Children | Wide Entrance | Peak Hour | Low Light | Overall |
|---|---|---|---|---|---|---|---|
| System A (Stereo AI) | 98.2% | 94.1% | 93.8% | 91.5% | 95.7% | 89.4% | 96.4% |
| System B (Mono AI) | 97.1% | 90.3% | 88.5% | 86.2% | 92.8% | 85.1% | 93.1% |
| System C (Thermal) | 95.8% | 78.4% | 82.1% | 84.7% | 88.3% | 94.6% | 88.9% |
| System D (Infrared) | 94.2% | 72.6% | 68.4% | 79.1% | 81.5% | 91.8% | 83.2% |
| System E (Hybrid) | 98.5% | 95.8% | 94.2% | 93.1% | 96.2% | 92.7% | 97.1% |
Overall Accuracy by People Counting System
- System A (Stereo AI) — accuracy: 96.4
- System B (Mono AI) — accuracy: 93.1
- System C (Thermal) — accuracy: 88.9
- System D (Infrared) — accuracy: 83.2
- System E (Hybrid) — accuracy: 97.1
The hybrid AI + thermal system (System E) achieved the highest overall accuracy at 97.1%, narrowly edging out the stereoscopic AI system (96.4%). The performance gap widened significantly in challenging scenarios — group entries and children were the most differentiating tests. Traditional infrared beam counters showed the steepest accuracy decline in complex scenarios, dropping from 94.2% in single entry to just 68.4% with children and strollers.
What the Data Tells Us
Three key patterns emerged from our testing. First, AI-based people counting software consistently outperformed sensor-only solutions, particularly in complex scenarios involving groups and children. The ability to use computer vision to distinguish individual people — even when they're overlapping in the camera's field of view — gives AI systems a fundamental advantage that infrared and basic thermal sensors cannot match.
Second, vendor accuracy claims almost always reflect best-case performance. The average gap between claimed and observed accuracy was 6.2 percentage points. System D (infrared) showed the largest gap at 11.8 points, while System E (hybrid) showed the smallest at 1.9 points. This underscores the importance of on-site validation before committing to a people counting system.
Third, environmental factors matter more than most buyers realize. Low-light conditions affected AI video systems more than thermal or infrared alternatives, though newer AI models with infrared-enhanced cameras are closing this gap. Wide entrances remain challenging for all technologies, suggesting that entrance design should be a key factor in system selection.
Our Recommendations
For retailers seeking the best people counting software, AI-based systems offer the highest accuracy in real-world conditions. If budget is a primary concern, monocular AI video (System B profile) offers a strong balance of accuracy and affordability. For environments with challenging lighting or extreme accuracy requirements, hybrid AI + thermal systems provide the most consistent performance across all scenarios.
We recommend every buyer conduct their own on-site accuracy test before purchasing. Request a trial period of at least 2 weeks, and manually verify counts during your busiest periods. The scenarios where accuracy drops most are exactly the times when accurate data matters most for business decisions.
Infrared beam counters remain viable for narrow, single-file entrances with consistent lighting — but should not be relied upon for multi-lane entrances or locations with significant group traffic. Their lower cost is quickly offset by unreliable data that can lead to poor staffing and marketing decisions.