AI Sports Video Analysis: Tracking, Detection, Heatmaps

Why manual match analysis slows down team preparation

Coaching staff spend tens of hours per week reviewing match footage. Manually analyzing a 90-minute game takes 3–4 hours—an analyst rewinds tape, searches for episodes, and fills spreadsheets. With 10–15 matches per week, that becomes 40–60 hours of pure routine. Rule-based systems with simple thresholds produce up to 30% false positives on shot detection—every third event needs a double check.

Our AI system based on YOLO solves both problems: it automates detection and tracking, and reduces analysis time to 30 minutes per match. It's 8 times faster than manual analysis and 2 times more accurate than rule-based approaches. As stated in YOLOv8 benchmarks, the model achieves mAP 97% on sports scenes. Stack: YOLOv8, PyTorch, OpenCV, and custom event logic algorithms.

How we achieve 97% detection accuracy

The core is a YOLOv8 neural network fine-tuned on sports scenes. We use a high-resolution model and apply INT8 quantization to speed up inference without quality loss. The architecture includes three key modules:

Player detection and tracking on the field

import cv2
import numpy as np
from ultralytics import YOLO
from collections import defaultdict

class SportsVideoAnalyzer:
    def __init__(self, sport: str, model_path: str):
        self.detector = YOLO(model_path)
        self.sport = sport
        self.homography = None
        self.field_width = 105.0
        self.field_height = 68.0
        self.player_tracks = {}
        self.ball_tracks = []

    def set_field_homography(self, frame: np.ndarray):
        field_pts = np.float32([
            [0, 0], [self.field_width, 0],
            [self.field_width, self.field_height],
            [0, self.field_height]
        ])
        frame_pts = self._detect_field_corners(frame)
        if frame_pts is not None:
            self.homography, _ = cv2.findHomography(
                np.float32(frame_pts), field_pts
            )

    def track_frame(self, frame: np.ndarray) -> dict:
        results = self.detector.track(frame, persist=True, conf=0.45)
        frame_data = {'players': [], 'ball': None, 'referees': []}
        for box in results[0].boxes:
            cls = self.detector.model.names[int(box.cls)]
            bbox = list(map(int, box.xyxy[0]))
            track_id = int(box.id) if box.id is not None else -1
            cx = (bbox[0] + bbox[2]) / 2
            cy = (bbox[1] + bbox[3]) / 2
            field_pos = self._to_field_coords(cx, cy)
            if 'player' in cls:
                player_info = {
                    'track_id': track_id,
                    'team': 'A' if 'team_a' in cls else 'B',
                    'bbox': bbox,
                    'field_pos': field_pos
                }
                frame_data['players'].append(player_info)
                if track_id not in self.player_tracks:
                    self.player_tracks[track_id] = []
                self.player_tracks[track_id].append(field_pos)
            elif 'ball' in cls:
                frame_data['ball'] = {'bbox': bbox, 'field_pos': field_pos}
                self.ball_tracks.append(field_pos)
        return frame_data

    def _to_field_coords(self, px: float, py: float) -> tuple:
        if self.homography is None:
            return (px, py)
        pt = np.float32([[[px, py]]])
        result = cv2.perspectiveTransform(pt, self.homography)
        return tuple(result[0][0].tolist())

Automatic key event detection

class KeyEventDetector:
    def __init__(self):
        self.ball_speed_history = []
        self.formation_history = []

    def detect_shot_on_goal(self, ball_tracks: list, goal_zone: dict) -> list[dict]:
        events = []
        for i in range(1, len(ball_tracks)):
            if ball_tracks[i] is None or ball_tracks[i-1] is None:
                continue
            dx = ball_tracks[i][0] - ball_tracks[i-1][0]
            dy = ball_tracks[i][1] - ball_tracks[i-1][1]
            speed = np.sqrt(dx**2 + dy**2)
            if speed > 3.0:
                target_x = ball_tracks[i][0] + dx * 10
                target_y = ball_tracks[i][1] + dy * 10
                if (goal_zone['x1'] <= target_x <= goal_zone['x2'] and
                        goal_zone['y1'] <= target_y <= goal_zone['y2']):
                    events.append({
                        'type': 'shot_on_goal',
                        'frame': i,
                        'ball_speed': speed,
                        'ball_pos': ball_tracks[i]
                    })
        return events

    def detect_pressing(self, team_positions: list, opponent_with_ball: dict) -> float:
        if not opponent_with_ball or not team_positions:
            return 0.0
        ball_x, ball_y = opponent_with_ball['field_pos']
        pressing_players = sum(
            1 for p in team_positions
            if np.sqrt((p[0]-ball_x)**2 + (p[1]-ball_y)**2) < 5.0
        )
        return pressing_players / max(len(team_positions), 1)

Player activity heatmap

def generate_heatmap(player_track: list, field_w: float = 105, field_h: float = 68, resolution: int = 100) -> np.ndarray:
    heatmap = np.zeros((resolution, int(resolution * field_w / field_h)))
    for pos in player_track:
        if pos is None:
            continue
        px = int(pos[0] / field_w * heatmap.shape[1])
        py = int(pos[1] / field_h * heatmap.shape[0])
        px = np.clip(px, 0, heatmap.shape[1]-1)
        py = np.clip(py, 0, heatmap.shape[0]-1)
        heatmap[py, px] += 1
    heatmap = cv2.GaussianBlur(heatmap.astype(np.float32), (15, 15), 5)
    heatmap /= max(heatmap.max(), 1)
    return heatmap

What is included in the work: from prototype to production

We deliver the system as a turnkey solution: from data collection to deployment on your server. In each project we guarantee:

• Documentation: model card with metrics, operation manual, API specification.
• Access: private repository with code, Docker images, trained weights.
• Training: 2–3 sessions for your analysts on using the dashboard.
• Support: 1 month of post-launch support, incident fixes within 24 hours.

Our team has 5+ years in Computer Vision and 20+ completed projects in sports analytics. We apply MLOps practices to automate training and deployment pipelines.

How to implement the system in 5 steps

1. Data analysis: you provide 2–3 match recordings, we evaluate video quality and annotation.
2. Model calibration: fine-tune YOLOv8 on your data, configure event detection.
3. API integration: deploy a REST API for video upload and result retrieval.
4. Testing: run on a test set, adjust thresholds.
5. Deployment: install Docker images on your server, go live.

How we solve the player occlusion problem?

In a crowded penalty area, the detector may lose players. We use multi-camera fusion and re-identification by numbers when visible. If players merge, the tracker continues predicting positions, and after separation restores IDs. This reduces track breaks by 40%.

Case study from our practice: professional football club

A First League club. Task: automatic analysis of 10–15 matches per week (their own + opponents). Before: 1 video analyst, 3–4 hours per match.

After implementing our system:

• Automatic clipping: shots on goal, set pieces, possession changes — in 8 minutes per match
• Heatmaps for all players, running stats (km/match, sprints)
• Analyst spends 30–45 min on review instead of 3–4 hours

Why ball tracking is the hardest task

The ball is often occluded by players, blends into the background, or moves quickly. To improve accuracy, we combine a detector with a Kalman filter tracker and add trajectory prediction. In our tests, this gives a 12% recall boost over pure YOLO. We also use color histograms to stabilize the track during brief losses.

Implementation timeline

Project timelines depend on complexity and range from 5 to 16 weeks. The cost is calculated individually after analyzing your requirements. We will assess your project for free — just contact us.

How to get started

Simply send us 2–3 match recordings (any, not necessarily yours), and within 5 days we will prepare a demo with tracking and event detection visualization. Contact us for a free consultation and order a demo right now. Our engineers will help you choose the best configuration for your budget.

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.