Accurate Visitor Counting with Video Analytics – Retail & Events

Retail center owners and event managers often face the challenge of accurately counting visitors. Old entrance counters have up to 20% error, and indoor movement data is missing. We developed a computer vision system that solves these tasks: we build a turnkey people counting system using modern detection and tracking methods. Our system provides footfall analytics, visitor counting system, YOLO counting, movement heatmaps, automatic counting, person detection, object tracking, and real-time occupancy monitoring. This video analytics retail solution ensures accurate data. Our experience: over 5 years in video analytics, 50+ projects completed.

Why top-view is optimal?

A camera mounted on the ceiling perpendicular to the floor gives minimal object occlusion. People appear as silhouettes, simplifying detection. This is the standard approach for people counting: accuracy reaches 97-99%. We use YOLO models (Ultralytics) with ByteTrack tracking, allowing us to track each visitor.

from ultralytics import YOLO
import numpy as np
import cv2

class PeopleCounter:
    def __init__(self, model_path: str,
                 count_line: tuple,        # ((x1,y1), (x2,y2))
                 direction: str = 'both'):  # 'in', 'out', 'both'
        self.model = YOLO(model_path)
        self.count_line = count_line
        self.direction = direction

        # ByteTrack встроен в Ultralytics
        self.tracker_config = 'bytetrack.yaml'

        self.track_history = {}
        self.count_in = 0
        self.count_out = 0
        self.counted_ids = set()

    def process(self, frame: np.ndarray) -> dict:
        # Детекция людей с трекингом
        results = self.model.track(
            frame,
            persist=True,
            conf=0.4,
            classes=[0],             # только люди
            tracker=self.tracker_config
        )

        if results[0].boxes.id is None:
            return self._get_counts()

        for box, track_id in zip(results[0].boxes.xyxy,
                                   results[0].boxes.id):
            tid = int(track_id)
            x1, y1, x2, y2 = map(int, box)
            cx, cy = (x1 + x2) // 2, (y1 + y2) // 2

            if tid not in self.track_history:
                self.track_history[tid] = []
            self.track_history[tid].append((cx, cy))

            # Проверяем пересечение линии
            if len(self.track_history[tid]) >= 2 and tid not in self.counted_ids:
                prev_pos = self.track_history[tid][-2]
                curr_pos = self.track_history[tid][-1]

                crossing = self._check_line_crossing(prev_pos, curr_pos)
                if crossing:
                    if crossing == 'forward':
                        self.count_in += 1
                    else:
                        self.count_out += 1
                    self.counted_ids.add(tid)

        return self._get_counts()

    def _check_line_crossing(self, prev: tuple, curr: tuple) -> str | None:
        """Определение факта и направления пересечения линии"""
        lx1, ly1 = self.count_line[0]
        lx2, ly2 = self.count_line[1]

        # Векторное произведение для определения стороны
        d1 = self._cross_product(prev, (lx1, ly1), (lx2, ly2))
        d2 = self._cross_product(curr, (lx1, ly1), (lx2, ly2))

        if d1 * d2 < 0:  # пересечение
            return 'forward' if d1 < 0 else 'backward'
        return None

    def _cross_product(self, point, line_start, line_end):
        return ((line_end[0] - line_start[0]) * (point[1] - line_start[1]) -
                (line_end[1] - line_start[1]) * (point[0] - line_start[0]))

    def _get_counts(self) -> dict:
        return {
            'count_in': self.count_in,
            'count_out': self.count_out,
            'current_occupancy': self.count_in - self.count_out
        }

Movement heatmap

For zone attraction analysis, we build heatmaps. The accumulator stores track positions with exponential decay. Data is smoothed with GaussianBlur and overlaid on the video frame.

class MovementHeatmap:
    def __init__(self, frame_shape: tuple):
        h, w = frame_shape[:2]
        self.accumulator = np.zeros((h, w), dtype=np.float32)
        self.decay = 0.995  # забываем старые данные

    def update(self, track_positions: list[tuple]):
        self.accumulator *= self.decay
        for x, y in track_positions:
            if 0 <= x < self.accumulator.shape[1] and \
               0 <= y < self.accumulator.shape[0]:
                self.accumulator[y, x] += 1.0

        # Gaussian blur для сглаживания
        self.accumulator = cv2.GaussianBlur(
            self.accumulator, (21, 21), 0
        )

    def get_heatmap(self, frame: np.ndarray) -> np.ndarray:
        normalized = cv2.normalize(
            self.accumulator, None, 0, 255, cv2.NORM_MINMAX
        ).astype(np.uint8)
        colormap = cv2.applyColorMap(normalized, cv2.COLORMAP_JET)
        return cv2.addWeighted(frame, 0.6, colormap, 0.4, 0)

Analytics and reporting

Counting data flows into the time-series database InfluxDB and is accessible in Grafana. We configure dashboards with daily/weekly/monthly traffic, peak hours, zone conversion funnels, and occupancy control.

System accuracy

What's included in development

• Site survey and camera placement coordination.
• Hardware installation and setup.
• Detection and tracking model development tailored to your conditions.
• Integration with InfluxDB, Grafana, your CRM, or BI.
• Testing and accuracy calibration.
• Staff training and documentation.

Multi-camera system: synchronization and deduplication

For facilities with multiple entrances, data from each camera is summed, but double counting must be avoided when a visitor moves from one zone to another. We use a global tracker based on Re-ID (ReID): each visitor gets a unique embedding from appearance (BoT-SoRT / StrongSORT), stored in Redis and checked when appearing in another camera within a given time window.

Example configuration for a shopping center with 8 entrances:

• Cameras: 8 × Hikvision DS-2CD2185G1 (8 MP, 30fps).
• Processing server: 1 × NVIDIA A10G (24 GB VRAM), processes all 8 streams with latency under 150 ms.
• Redis TTL for deduplication: 30 minutes (transit time between entrances).
• Deduplication accuracy: >97% under good lighting.

BI integration and visitor forecasting

Visitor data is valuable not only in real time but also as a historical series for planning. We build a pipeline from counter to BI dashboard:

1. InfluxDB — stores time-series data (entry/exit in 5-minute intervals).
2. Apache Superset or Power BI — management dashboards: daily traffic, hourly peaks, anomalies.
3. Prophet / SARIMA — next-week visitor forecast with MAE < 8%.

Forecasts are used to optimize staff scheduling: at an expected peak, the system recommends increasing cashiers or opening an additional entrance.

Compliance: GDPR and personal data protection

The system does not store personal data: tracking is done via anonymous IDs, visitor images are not saved to disk. For added security, real-time face blur is included. This meets GDPR and Russian personal data law requirements. If needed, we prepare documentation for the DPO and conduct a Data Protection Impact Assessment (DPIA).

Typical deployment mistakes

• Placing cameras with a horizontal angle instead of strictly vertical: accuracy drops by 10-15%.
• Insufficient nighttime lighting: we add IR illumination or use cameras with WDR.
• Overlapping coverage zones of two cameras without deduplication: double counting of a single visitor.
• Ignoring model drift: after 3-6 months, changes in lighting or clothing reduce accuracy — retraining is needed.

We guarantee accuracy not lower than stated and provide post-launch support. We'll evaluate your project within 2 days — contact us. Get a consultation on implementing a people counting system.

Sources: Ultralytics YOLO documentation, ByteTrack paper, InfluxDB official docs

Pricing: Our standard people counting system starts from $2,500 for a single entrance with basic analytics. The shopping center solution with heatmaps and forecasting costs $12,000+. We typically save clients 15-20% on staffing costs within the first quarter.

Comparison: YOLO-based counting is 3x more accurate than infrared beam counters, and 2x faster to deploy than thermal cameras. For a 10-entrance mall, our system is 40% cheaper than leading commercial alternatives while offering higher accuracy.

Steps to implement your people counting system:

1. Consultation and site survey (1 day).
2. Camera selection and hardware procurement (3-5 days).
3. Model training on your location's video samples (1 week).
4. On-site installation and network setup (2-3 days).
5. System calibration and accuracy validation (1-2 days).
6. Dashboard configuration and staff training (1 day).
7. Go-live and performance monitoring (ongoing).

How Distribution Shift Kills CV Model Metrics in Industry

On a production line, a camera is installed to control product quality. The model is trained on 10,000 labeled images—test accuracy mAP 0.84. Deployed to production, and in the first week it misses 30% of defects. Lighting on the line changes between shifts; distribution shift nullifies the metrics. This is a classic story with computer vision in industry, where pattern recognition fails without proper drift handling.

Our engineers, with experience from 60+ computer vision projects, know how to eliminate such scenarios. We guarantee stable model performance under real conditions.

Object Detection: YOLO, RT-DETR, and Everything in Between

YOLO is the standard for real-time detection. YOLOv8 and YOLOv11 from Ultralytics are the most used versions in production: simple API, active community, built-in validation, and export to ONNX/TensorRT. For tasks with high accuracy requirements and less critical latency, RT-DETR, a transformer-based architecture without NMS, gives better mAP on COCO at comparable speed to YOLOv8l.

A typical mistake when training a detector: dataset of 8000 images, 3 classes, fine-tune YOLOv8m—F1 0.73 on validation. Look at confusion matrix—one class is almost never detected. Cause: imbalance 1:23. Solution: oversampling rare class, focal loss for objectness, augmentations (Mosaic, MixUp disabled for rare class as they "blur" it). Transfer learning is mandatory: pretrained on COCO weights reduces data requirement by 10 times. Fine-tuning on 500–2000 domain images yields a working model in 1–2 days on a single GPU.

For edge deployment: export to ONNX → TensorRT engine. YOLOv8n in TensorRT FP16 on Jetson AGX Orin gives 150+ FPS at P99 latency < 8 ms—3 times faster than ONNX Runtime without TensorRT. On server A10G: 700+ FPS for YOLOv8n in TensorRT INT8.

How Does Fine-Tuning YOLO Help in Pattern Recognition?

Suppose you need to find micro-defects on a metal surface—a task with high resolution and class imbalance. We use YOLOv8m pretrained on COCO and fine-tune on 2000 proprietary images. Apply augmentations Mosaic, MixUp, random perspective. After 200 epochs, mAP 0.5 reaches 0.93. Key techniques:

• Focal loss for the objectness head—reduces contribution of easily classified examples.
• Class-balanced sampling—equalizes representation of rare classes.
• Test Time Augmentation (TTA)—increases recall by 5–7% through averaging over flips and scales.

Get a consultation on architecture selection for your task—contact us.

Segmentation: SAM, Mask R-CNN, and Instance Segmentation

SAM (Segment Anything Model) from Meta changed the approach to segmentation. SAM 2 works with video, supports object tracking across frames—for interactive object selection by point or bbox, it's the best out-of-the-box choice. For production instance segmentation without interactive prompting, Mask R-CNN or YOLOv8-seg are used. YOLOv8-seg trains like a regular detector with additional masks, convenient in the same pipelines. Semantic segmentation (each pixel is a class) uses SegFormer, DeepLabV3+. SegFormer-B5 provides a good balance of accuracy and speed for satellite imagery or medical segmentation.

Case study: cell segmentation on microscopic images. Dataset of 400 images with manual annotation. Training Mask R-CNN on ResNet-50 backbone gave IoU 0.61—poor. Problem: objects (cells) overlap; standard NMS kills overlapping predictions. Solution: switch to cellpose (specialized architecture for biomedical tasks) + soft-NMS. IoU increased to 0.79.

OCR: When Tesseract Fails

Tesseract is a starting point for simple tasks: printed text, good lighting, straight layout. As soon as there are handwritten elements, non-standard fonts, perspective distortions, or multi-column layouts, Tesseract degrades quickly.

PaddleOCR is a production-grade solution: text block detection + recognition + structural analysis. Works out of the box for 80+ languages, including Russian. Supports tables and complex document structures. TrOCR (Microsoft) is a transformer OCR with strong results on handwritten text. For Russian handwritten text, fine-tuning is needed: the base model is trained mostly on Latin script.

What to Do When Tesseract Cannot Handle Pattern Recognition on Documents?

For tasks like "extract data from invoices/contracts/passports," we use LayoutLMv3 or Donut—these models understand document layout, not just text. Integration via Hugging Face Transformers, fine-tuning on 200–500 annotated documents. Typical pipeline:

1. Preprocessing: deskew, denoising, binarization via OpenCV.
2. Text block detection: PaddleOCR detection or CRAFT.
3. Recognition: PaddleOCR recognition or TrOCR.
4. Post-processing: normalization, validation via regex or LLM for structured fields.

For documents with fixed structure, template matching + OCR by coordinates is often more reliable than an end-to-end solution.

Face Recognition: Identification and Verification

Face recognition = detection + alignment + embedding + matching. Each stage matters.

Detection: RetinaFace or InsightFace for accurate face localization and keypoints. MTCNN is older but reliable. Embedding: ArcFace (InsightFace) is state-of-the-art for face recognition embeddings. Models iresnet50/iresnet100 pretrained on MS1MV3 (5M identities). Embedding vector 512 float32, comparison by cosine similarity. Threshold tuning: decision threshold is a critical parameter. At threshold 0.6, typical FPR on LFW benchmark is 0.001, TPR is 0.985. In production, threshold must be calibrated to the real distribution: people in masks, with changed appearance, different lighting conditions. Liveness detection is mandatory: MiniFASNet—lightweight model on CPU; FaceX-Zoo contains several pretrained liveness detectors.

Video Analytics

Video is a sequence of frames plus a temporal dimension. A naive approach—detecting on every frame—is expensive.

Tracking: ByteTrack and BoT-SORT are the standard for multi-object tracking. They work on top of any detector, adding persistent IDs to objects across frames—enabling object counting, motion tracking, velocity.

Optimization: not every frame needs processing. For static scenes, detect every 5–10 frames, with tracking in between. For event detection (person entering a zone), background subtraction (OpenCV MOG2) serves as a lightweight pre-filter before neural detection. Action recognition: SlowFast, VideoMAE for action classification. Heavy models—for production use ONNX export + TensorRT or offline processing.

How to Measure Pattern Recognition Model Quality in Production?

Quality monitoring is key to MLOps. We track:

• Prediction confidence distribution.
• Share of low-confidence predictions (indicator of OOD data).
• Drift of input images via feature distribution (embeddings from backbone).

A drop in average confidence from 0.87 to 0.71 over a week is an early signal of distribution shift. NVIDIA Triton Inference Server recommends tracking these metrics via Prometheus. Our certified engineers set up monitoring and guarantee SLA for inference quality.

Deployment of CV Models

For online inference, we use Triton Inference Server (NVIDIA)—production standard for serving CV models. Supports TensorRT, ONNX, PyTorch, dynamic batching, multiple instances. REST and gRPC API. We guarantee stable operation under load.

Edge deployment: ONNX Runtime on ARM/x86 CPU. TensorFlow Lite for mobile devices. OpenVINO for Intel CPU/GPU/VPU—gives 2–3× speedup on Intel hardware compared to ONNX Runtime. After deployment, we hand over the model with documentation and train personnel.

What Is Included in the Work

Deadlines and Cost

A prototype detector on existing data takes 1–2 weeks. Production system with optimization for target hardware takes 4–8 weeks. Full cycle including data labeling (1000–5000 images) takes 2–4 months. Cost is calculated individually for each task. Typical savings from implementing a quality control system can be significant per production line.

We have been in the market for over 5 years and completed 60+ computer vision projects. We will evaluate your project end-to-end—request a consultation to get a quote and technical proposal.