Computer Vision Developer – Sports Player Tracking & XY Coordinates POC

@sajjadtaghvaeifr

Hello, I trust you're doing well. I am well experienced in machine learning algorithms, with nearly a decade of hands-on practice. My expertise lies in developing various artificial intelligence algorithms, including the one you require, using Python, and similar tools. I have worked with pytorch, and tensorflow to develop DL models, .I hold a doctorate from Tohoku University and have a number of publications in the same subject. My portfolio, which showcases my past work, is available for your review. Your project piqued my interest, and I would be delighted to be part of it. Let's connect to discuss in detail. Warm regards. please check my portfolio link: https://www.freelancer.com/u/sajjadtaghvaeifr

@ShaikhAneesa

Hi there, I’ve carefully reviewed your requirements and understand that the real test here is not simply detecting players, but reliably maintaining who is who across broadcast footage, mapping those players to roster identities and field coordinates, and knowing when the evidence is not strong enough to make an identification. I’m confident I can build a focused POC that demonstrates this properly while keeping the architecture flexible for multiple sports and future live video. My approach will start by analyzing the footage, rosters and camera views to establish detection and calibration constraints. Then, I’ll use Python, OpenCV and YOLO for player, team and ball detection, with ByteTrack or BoT-SORT for tracking and ReID embeddings for identity continuity. Next, I’ll implement roster-assisted matching, jersey-number recognition where visible, homography calibration and real-world x/y coordinates. Finally, I’ll apply confidence thresholds, preserve UNKNOWN when evidence is insufficient, and export timestamped CSV/JSON results. You will receive the working POC, tracking and ReID pipeline, roster identification, calibration, coordinate output, CSV/JSON export and feasibility findings. Could you share the sample footage and rosters so I can assess the camera angles, player visibility and field markings? I’m ready to prove the core feasibility with measurable, defensible results rather than forcing unreliable detections. Warm Regards, Aneesa.

@ZohaibRoy

⭐⭐⭐⭐⭐ Create a Multi-Sport Video Analysis System with Computer Vision ❇️ Hi My Friend, I hope you're doing well. I've reviewed your project needs and noticed you're looking for an experienced computer vision developer. You don't need to look any further; Zohaib is here to help you! My team has successfully completed 50+ similar projects in video analysis. I will employ efficient methods to ensure player detection, tracking, and data output, all within your budget. ➡️ Why Me? I can easily build your multi-sport video analysis system as I have 5 years of experience in computer vision, specializing in player detection, tracking, and data processing. My expertise includes working with video footage, real-time tracking, and data output formats. Additionally, I have a strong grip on technologies like OpenCV and machine learning, which will enhance the project's success. ➡️ Let's have a quick chat to discuss your project in detail and let me show you samples of my previous work. I look forward to discussing this with you soon! ➡️ Skills & Experience: ✅ Computer Vision ✅ Player Detection ✅ Video Analysis ✅ Machine Learning ✅ Continuous Tracking ✅ Data Output (CSV/JSON) ✅ Field Calibration ✅ Homography ✅ Real-time Processing ✅ Python Programming ✅ OpenCV ✅ Sports Analytics Waiting for your response! Best Regards, Zohaib

@AlonzoLeeeooo

For a sports tracking POC the hard parts are identity continuity through occlusions and honest coordinates, not the detector itself. I've worked on tracking and re-identification in video, and I know how fast track IDs fragment when players cluster or the camera pans; my approach pairs a strong detector with a ReID step that matches on appearance and keeps tracklets alive through short occlusions instead of reassigning IDs. Jersey-number recognition slots in as a disambiguation signal on top of that, and for the x/y output I'd calibrate each clip with a homography from field markings, then flag frames where the calibration is uncertain rather than silently outputting bad coordinates. I'm a USTC Ph.D. with 11 published papers across computer vision and generative models, so detection and tracking are squarely in my area. Plan: (1) detector plus tracking and ReID baseline on sample footage; (2) roster-assisted named identification and jersey-number recognition; (3) pitch calibration and real-world x/y export as timestamped CSV/JSON; (4) ball detection and a short accuracy note with failure cases. One question to start: which sports are in the sample footage, and is the camera fixed or broadcast-style? https://www.freelancer.com/u/AlonzoLeeeooo

@bojans9

Hi, Bojan here, from Serbia. Building your computer vision comes down to one thing: answers people can trust, at a running cost that stays predictable. I'm a senior AI developer with 8+ years of experience working with Computer Vision, Machine Learning (ML) and Deep Learning. For your computer vision, the real challenge is wrong answers and prompt injection. A common issue: launching without guardrails or a test set, so nobody knows how often it is wrong. My plan: first, collect real examples and write down the expected result for each. Then, build the LLM pipeline with guardrails and logging. Finally, score it against those examples and tune it before launch. I'd start with the core assistant and its test set as phase one. I've done AI work before, so I check early for wrong answers and prompt injection. A few questions to better understand the scope: 1. Must every action be logged for review? Full logs make every mistake traceable later. 2. Hand off to a person when it is unsure? I'd recommend it, handoffs beat wrong answers. 3. Who relies on its results, customers or your team? Output for customers needs stricter checks. I can start right away. Happy to go over the details here in chat and suggest the best approach. Looking forward to hearing from you, Bojan

@ImranAli01

Hello, I went through your project description and understand you need a focused computer vision POC to test player detection, roster-assisted identification, tracking/ReID, jersey numbers, field calibration, player coordinates, and ball tracking from broadcast sports footage. I will build the POC in Python using OpenCV, PyTorch, and suitable object-detection/tracking models such as YOLO with ByteTrack or BoT-SORT. Roster data will support identity association, while OCR will be applied for jersey numbers where image quality permits. Homography calibration will convert image coordinates into real-world field coordinates, and the pipeline will preserve UNKNOWN whenever detection, tracking, or identity confidence is insufficient. Timestamped CSV/JSON output will provide player IDs, teams, positions, confidence, and ball data. The architecture will remain modular so additional sports and future live-stream processing can be supported. I will share a project blueprint within 24–48 hours of being awarded the project, covering the CV pipeline, tracking strategy, calibration, outputs, and POC milestones. Could you share the sample footage and roster format so I can align the POC with the actual camera angles and data? Let’s validate the core tracking pipeline first. Best Regards, Imran

@FreeLancers360

Hi there, Your project "Computer Vision Developer – Sports Player Tracking & XY Coordinates POC" instantly caught my attention — it's exactly the kind of work we love to take on. I love turning a solid brief like yours into a polished result you’ll be proud to put your name on. From your brief I can see this involves ai, machine learning, computer vision — all areas we handle in-house. We specialise in Machine Learning (ML), Image Processing, Computer Vision, Deep Learning, which lines up directly with what you need. How we'd approach it: - Clarify the use-case, inputs and the exact output you expect - Build and integrate the model / automation pipeline - Evaluate accuracy and tune against real examples - Deploy with monitoring and a clear handover I’ll fold in your feedback fast and keep refining until the result feels exactly right to you. If it helps, I can share a couple of relevant samples and a short plan before you decide. Looking forward to it! Best regards, FreeLancers360 Let’s connect in chat and get started — message me anytime and I’ll reply right away!

@adriank06

I can build this POC. My pipeline: YOLOv8/RT-DETR for player/ball detection, BoT-SORT/DeepSORT for tracking & ReID, jersey OCR with EasyOCR/PaddleOCR, homography from field markings to real-world x/y, and timestamped CSV/JSON with confidence scores. Low-confidence identities or positions stay UNKNOWN – no forced guesses. Architecture will be sport-agnostic and ready for later live-stream operation. I’ll process your sample footage, document failure cases (occlusion, camera cuts), and deliver a reproducible POC with sample outputs. Quick question: What’s the sample clip length and camera angle – fixed broadcast or panning/zooming? Also, do you have roster photos or just names/numbers? This drives ReID and jersey OCR accuracy. I have 5 years in software and 2+ years in AI/CV, with hands-on experience in detection, tracking, and spatial mapping. I can share relevant CV work in chat. Available to start immediately.

@devendrathakur12

i’ve done very similar recently with OpenCV, YOLO, tracking/ReID, OCR, homography, and Python video pipelines. I’d keep identity separate from detection so uncertain cases remain UNKNOWN instead of producing misleading player labels. Do you have camera calibration points or should the POC estimate field homography from visible markings? Are the supplied rosters structured with known jersey numbers and player metadata? I’d use ByteTrack/BoT-SORT with ReID features and confidence gating, which improves continuity through short occlusions. I’d also keep sport-specific rules outside the core tracker so the same pipeline can later support live streams. I’ll first establish detection and calibration, then add tracking, roster matching, jersey OCR, and ball detection. Finally, I’ll produce timestamped JSON/CSV with confidence and UNKNOWN states, plus clear POC results and failure cases. Best, Dev S.

@shadabkhan92

With over X years of experience in deploying and implementing computer vision systems across various industries, my team and I are the perfect fit for your sports player tracking project. Our AI expertise goes beyond prototypes -- we focus on building production-ready solutions that seamlessly integrate into your existing workflows. We can assure you a reliable and high-performing system that not only tracks players but also empowers you with visual insights for strategic decision-making. To ensure quality results, our technical approach entails a combination of robust machine learning algorithms, smart calibration methods, and advanced object detection techniques. This, complemented with our in-house built hardware and IoT proficiency, means we can provide a comprehensive solution- from real-time feed processing to x/y-coordinate output. Moreover, having worked extensively on multi-sport applications before, we understand the challenges revolving around genre agnostic designs. Thus, our architecture will be designed in a way that allows future opportunity for transition to live-stream operations without limiting you to any particular sport. Lastly, we are familiar with delivering timestamped CSV/JSON outputs to make data utilization more convenient for you. Choose us for this POC not just for a potential commercial application but also to establish a solid partnership opening doors to broader developments!

@aakaashberi

Keeping UNKNOWN when identity or position isn't reliable is the right call. A POC that guesses would hide exactly the gaps you need to see before scaling this. Approach: a YOLO detector for players and the ball, ByteTrack plus a ReID embedding for continuous tracks, jersey OCR on crops to fuse with the roster, and pitch-keypoint homography to turn pixels into metres. Each stage is a separate module with confidence scores, so it stays sport-agnostic and can move to a live stream later. Deliverables: Python pipeline + README, timestamped CSV/JSON (frame, track id, player/UNKNOWN, x/y, confidence), an annotated overlay video and a short feasibility note on what worked and where accuracy drops. First working version in 3 days. Price: AUD 250 for the POC. My past client work is under NDA, so I can't share it, but the overlay video on your own clip will be the proof. Questions: which sport and how long is the sample clip? And is the camera a single broadcast feed with cuts and zooms? Aakaash

@jeanpatrickr8

Hello, I’m excited about the opportunity to develop a proof of concept for your multi-sport video analysis system. I understand that you need a robust solution for player detection, identification, tracking, and data output using broadcast field-sports footage. I have extensive experience in computer vision, particularly in sports analytics, and have successfully implemented projects involving player tracking and identification using technologies like OpenCV and deep learning frameworks. My background equips me to handle the complexities of your project efficiently. To ensure a successful feasibility POC, I propose the following approach: - Analyze the provided footage to establish baseline detection and tracking algorithms. - Implement player and team detection using machine learning models tailored for sports environments. - Develop a robust system for roster-assisted identification and jersey number recognition, ensuring accurate output. - Design the architecture to support future live-stream capabilities, maintaining flexibility for various sports. I am eager to begin this project and confident in delivering quality results that meet your expectations. Let’s discuss any further details at your convenience, as I am ready to start immediately.

@C0RETECHIES

Hello, System understanding: The POC will ingest short broadcast clips, run player and team detection, cross‑reference supplied rosters for named‑player ID, maintain continuous tracks with ReID, attempt jersey‑number OCR, compute field homography to output real‑world X/Y coordinates per frame, and optionally track the ball. Unknown identities or positions are preserved as NULL rather than forced guesses, producing timestamped CSV/JSON. Technical approach: Use a detection backbone (YOLOv8 or Detectron2) for player/team segmentation, an appearance‑based ReID model (OSNet or similar) for track continuity, a lightweight OCR (Tesseract or CRNN) for jersey numbers where visible, and a homography step derived from field markings or known court dimensions. Outputs are written per‑frame to CSV/JSON with frame timestamp, player ID, team, X/Y, jersey number (if known), and ball coords. Core modules: 1) Video ingest & frame sampler; 2) Detector & team classifier; 3) ReID tracker with roster‑assisted ID assignment; 4) Jersey‑number recognition; 5) Field calibration & coordinate transform; 6) Ball detection/tracking (optional); 7) Result logger (CSV/JSON) with UNKNOWN handling. Relevant systems: - (no direct sports‑tracking CV projects in public portfolio; our computer‑vision and AI automation experience enables rapid prototyping). Implementation strategy: Begin with detector tuning on the supplied footage (2‑3 days), add ReID and roster mapping (2 days), implement homography and coordinate output (1‑2 days), integrate jersey OCR and ball tracking as time permits, validate with unknown‑handling, and deliver the POC package. Questions: 1. What is the expected length and frame rate of the broadcast sample, and are there multiple camera angles to consider? 2. Should the homography be derived from visible field lines, or will you provide calibration markers or known dimensions? 3. For jersey‑number recognition, is a confidence threshold acceptable, or must we output NULL when OCR is uncertain? Regards, Rohit

@ronakkataria1

Homography only holds for points on the grass, and that catches two items on your list. Players: map the box centre and every position is off, because the centre is about a metre up. From a camera 20 m high, a player 50 m away lands 2.4 m too deep, so the foot point is what gets mapped. The ball is worse. At 5 m up and the same distance it lands about 17 m off, and one camera can't see height. So ball XY is written when it's on the ground and UNKNOWN in flight, which fits your rule. Approach: a YOLO-style detector for players, referees and ball, an online tracker with appearance ReID, teams split by kit colour, and jersey numbers read when a back faces the camera, voted along the track and checked against the roster. Pitch keypoints give a fresh homography every frame, since broadcast cameras pan and zoom. Frames with too few landmarks, and replays, write UNKNOWN. The tracker only looks back, so the same code can run live later, and each sport's pitch is a config file. Output is timestamped CSV and JSON with a confidence on every name and position. Background: ML products for real-time object detection and tracking on AWS. Nothing sports-specific to show, so this POC would be the proof. My rating here is 5 out of 5. For the ball, is ground position plus UNKNOWN in the air enough here, or do you need its flight path? That takes a second camera or a physics fit. Ronak

@AITSoft

Hi, I understand you need a small paid POC that detects players and teams, links players with supplied rosters, tracks them over time, reads jersey numbers where possible, and converts field positions into real-world x/y coordinates. The system should keep UNKNOWN when the footage is unclear instead of guessing, while remaining flexible for different sports and future live streams. My approach would combine Computer Vision with Machine Learning (ML) for detection, tracking, ReID, jersey-number recognition, and ball tracking. I would use Image Processing to improve broadcast frames and field calibration, then return timestamped results in CSV or JSON with confidence values and UNKNOWN states. I will keep the design modular so the POC can be tested quickly without locking it to one sport or video format. I can also explain the main accuracy limits and next steps after testing your sample footage. Could you share the sample footage, roster format, target sports, and whether field calibration points or camera metadata are available? I can begin with the supplied video and roster data and provide a clear feasibility result with sample outputs.

@mayr81

I can help build a focused, technically sound proof of concept for multi-sport player tracking from broadcast footage. The POC would be structured around: • Player and team detection with confidence scoring • Roster-assisted identity matching, jersey-number OCR, and explicit UNKNOWN handling • Persistent tracking and ReID across occlusions and camera movement • Field calibration using homography to convert image coordinates into real-world x/y positions • Ball detection and tracking when visibility and resolution support it • Timestamped CSV or JSON output with confidence and identity metadata I would keep the pipeline sport-agnostic by separating detection, tracking, calibration, identity association, and export layers. The initial work would establish measurable feasibility on the supplied sample, document failure cases, and avoid forcing unreliable identities or coordinates. The design can then be extended toward live-stream processing without being tied to a single sport or camera setup. I can provide relevant computer vision and object-tracking examples, explain the model and pipeline choices, and deliver a reproducible POC with clear evaluation results.

@corpmember29

Hello, I HAVE BUILT SIMILAR COMPUTER VISION AND VIDEO ANALYSIS SOLUTIONS AND CAN DEVELOP THIS SPORTS TRACKING POC. I have 11+ years of development experience and can build the POC using Python, OpenCV and modern object detection/tracking models. The pipeline can cover player/team detection, roster-assisted identification, tracking/ReID, jersey number recognition, field calibration/homography, ball tracking and real-world x/y coordinates. I will design the system so uncertain detections remain UNKNOWN rather than generating unreliable identities or positions. Output can be provided as timestamped CSV/JSON with confidence and tracking data. For the POC, I would first process the supplied broadcast footage, establish field calibration, detect and track players/ball, then integrate roster and jersey information for identity matching. The architecture will remain modular so it can later support additional sports and live-stream processing. I can also share relevant computer vision/video tracking work and start with the paid feasibility test. Thanks, Christina

@usmansiddiqui001

Hello, I’m an AI developer with hands-on experience in Computer Vision, machine learning, Python, and building production-oriented AI systems. I can approach this as a focused feasibility POC rather than overbuilding the first version. 1. Detect players, teams, and the ball where visibility permits using modern object-detection/tracking models. 2. Combine roster information, jersey appearance, jersey numbers, and tracking history for named-player identification. 3. Implement continuous tracking/ReID while preserving `UNKNOWN` whenever confidence is insufficient. 4. Calibrate the playing field using homography and convert image coordinates into real-world X/Y positions. 5. Produce timestamped CSV/JSON containing detections, identity, confidence, tracking ID, and coordinates. 6. Keep the architecture sport-agnostic so additional sports and eventually live-stream processing can be supported. 7. Benchmark the POC against the supplied footage and clearly document what works reliably and where limitations remain. I’ve worked with Python-based AI/Computer Vision projects and can share relevant examples/source samples where appropriate. For the POC, I’d first evaluate the supplied footage to determine camera angles, visibility, field markings, and jersey-number quality before selecting the detection/tracking pipeline. Do you already have a preferred CV framework/model, or would you like me to recommend the stack based on the sample footage? Regards Usman

@alfatrade2

Preserving UNKNOWN instead of forcing a result is the right call, but it means the ReID model needs a confidence threshold tuned per camera angle — broadcast footage with frequent occlusion will push a lot of frames into that bucket unless the roster photos are recent and high-res. Do you have access to individual player headshots, or just team rosters as names/numbers? I've built FastAPI pipelines with staged data processing for structured output, so wiring detection, homography and timestamped CSV/JSON into one coherent flow fits how I usually work. How long is the sample clip?