Source-linked AI summary
Deep Learning Based Illegal Bowling Action Detection
Debopom Sutradhar, Niful Islam, Sudipto Mondal, Tasmima Hossain Jamim, Jubaer Muhammad Shufol, Swakkhar Shatabda
TL;DR
Real-time detection of illegal bowling actions is challenging because current sensor-based assessment takes place after matches. The paper proposes a computer-vision pipeline that compares the bowling-arm angle at shoulder and release frames, and reports 91.96% accuracy with 100% recall on its evaluation dataset. The authors note that broader testing is needed to establish robustness across bowlers and conditions.
Problem
Real-time detection of illegal bowling actions remains challenging, while sensor-based methods rely on lengthy post-match analysis.
Source: §1 · p. 1Method
The system detects shoulder and release frames, measures the change in bowling-arm angle, and flags a delivery when the difference exceeds 15°.
Source: §2.2 · p. 3Results
91.96% accuracy and 100% recall were achieved by the MediaPipe–RT-DETR pipeline.
Source: §5.3 · p. 10Takeaways & Limitations
All illegal bowling actions in the test set were identified, while false-positive cases could be checked through biomechanical assessment.
Source: §5.3 · p. 10The dataset excludes female, side-arm, and left-handed bowlers, limiting the diversity represented in the evaluation.
Source: §6 · p. 11
Abstract
from arXiv · showhide
Cricket, often referred to as the "gentleman's game," adheres to a strict rule set for both batsmen and bowlers, where each delivery can significantly impact the match outcome. Detecting illegal bowling actions is crucial for maintaining fair play, yet it remains challenging for umpires to monitor in real time. Existing sensor-based solutions have limitations in live match scenarios, making real-time assessment difficult. This paper proposes a computer vision-based deep learning solution to detect illegal bowling actions in live cricket matches. To develop and evaluate our approach, we compiled a dataset of 62 videos featuring 11 male bowlers, capturing both legal and illegal bowling actions from multiple angles-front, back, and side. However, the dataset predominantly comprises right-handed bowlers with conventional actions. The proposed system identifies two key frames, the shoulder frame and the release frame from video footage of a bowler's delivery and analyzes the change in the bowling arm's angle between these frames. If the angle difference exceeds a predefined threshold (e.g., 15 degrees), the delivery is flagged as potentially illegal. We evaluated the system on a custom dataset and achieved a high true positive rate, suggesting the system's potential effectiveness in real-time match settings. However, further research is required to validate the system across diverse environmental conditions and larger datasets to ensure generalizability and robustness in various live match scenarios. To the best of our knowledge, this is the first AI-based computer vision method for detecting illegal bowling actions in cricket.
1 Introduction
Real-time detection of illegal bowling actions remains difficult because existing sensor-based methods rely on lengthy post-match analysis, while computer vision work has not addressed live-match detection. This paper proposes a computer-vision system and evaluates it on a labeled dataset of 62 videos.
1 Introduction
- Real-time illegal-action detection is an unmet computer-vision task, while current sensor-based methods require a 21-day post-match analysis.Related computer-vision applications in cricket include no-ball detection, decision-review support, and delivery-length estimation.Source: §1 · pp. 1–2
- The system compares bowling-arm angles at the shoulder and release frames, flagging deliveries when the difference exceeds 15°.MediaPipe identifies the shoulder frame and RT-DETR extracts the release frame.Source: §1 · p. 2
- The authors describe their approach as the first deep-learning method for automatically detecting illegal bowling actions in live cricket matches.They also made the dataset and source code publicly available.Source: §1 · p. 2
- The evaluation uses a first labeled dataset of legal and illegal bowling actions comprising 62 videos from 11 bowlers.The dataset supports evaluation of the proposed solution.Source: §1 · p. 2
- 91.96% accuracy and 100% true positive rate were reported for the proposed solution on a controlled dataset.The paper presents these results as evidence of potential effectiveness.Source: §1 · p. 2
2 Preliminaries
The preliminaries introduce the ICC’s 15° rule for identifying chucking, explain pose estimation, and describe object detection as a computer-vision technique used in the proposed pipeline.
2.1 ICC Rules for Bowling Action
- A delivery is considered chucking when the elbow-angle difference between shoulder level and release exceeds 15°.The existing assessment process includes post-match biomechanical analysis in an approved laboratory, taking 21 days to interpret and report.Source: §2.1 · pp. 2–3
2.2 Pose Estimation
- Pose estimation infers a person’s pose from images or video by locating joint landmarks and connecting them into a simplified skeleton.In this work, MediaPipe detects the shoulder and release frames and measures arm-angle difference against the 15° threshold.Source: §2.2 · p. 3
2.3 Object Detection
- Object detection locates objects in images or video; YOLO uses a single neural network to predict bounding boxes and class probabilities.Source: §2.3 · p. 3
3 Related Work
Prior cricket computer-vision work covers delivery and umpiring tasks, but the cited literature does not address live-match detection of illegal bowling actions without extra hardware.
3.1 Video based methods
- Video-based no-ball detection selects a base frame, identifies the popping crease, and compares pixel changes across the lawful and no-ball regions.Source: §3.1 · p. 4
3.2 Vision based umpiring
- Vision-based cricket work includes waist-high no-ball detection and support for umpiring decisions, as well as shot classification and delivery-length analysis.The cited examples use InceptionV3, ball tracking and regression, image features, YOLO, and RetinaNet.Source: §3.2 · p. 4
3.3 Automated Pitch Analysis
- Automated pitch-analysis work combines player and ball localization with delivery categorization, scoring, and crease-based no-ball decisions.The described approaches also include Hawk Eye evaluation of bowling turn and image-processing methods.Source: §3.3 · p. 4
3.4 Comparative methods
- Existing methods for detecting illegal bowling actions are not computer-vision approaches, and some require wearable sensors that are unsuitable for real-match scenarios.The proposed work addresses this gap with a computer-vision solution that requires no extra hardware.Source: §3.4 · p. 5
4 Proposed System
The system combines dataset preparation, shoulder- and release-frame extraction, and a two-frame angle comparison to classify bowling actions.
4.1 Dataset
- The dataset contains 62 videos from 11 bowlers, with 36 expert-labeled videos and 12 labeled instances reserved for testing with no test bowler in training.Videos are cropped to retain bowling moments; footage was collected from front, back, and side angles, with side view used first to assess legality.Source: §4.1 · p. 5
4.2 Shoulder Frame Extraction
- The shoulder frame is the frame whose hand-to-shoulder angle is closest to 270°, measured using Mediapipe.The system checks each frame and selects the one with the smallest absolute deviation from the expected angle.Source: §4.2 · p. 5
4.3 Release Frame Extraction
- A YOLOv8 model trained on manually annotated release frames returns the frame with the highest probability of showing the ball release.Roboflow was used to select release frames for training and validation; the best-performing validation checkpoint is retained.Source: §4.3 · p. 6
4.4 Bowling Action Detection
- A delivery is marked illegal when the measured elbow-flexion difference between the shoulder and release frames exceeds 15°; otherwise, it is considered legal.The decision uses only those two extracted frames and follows the stated ICC threshold.Source: §4.4 · p. 6
5 Experimental Results
The MediaPipe–RT-DETR pipeline achieved the section’s strongest illegal-action classification results, while release-frame localization and pose-estimation accuracy remained central to performance. The authors identify broader validation across bowling styles and environments as necessary.
5.1 Evaluation metrics
- The evaluation uses seven metrics: accuracy, precision, recall, F1-score, true positive rate, false positive rate, and MCC.The section presents them as different perspectives on model performance.Source: §5.1 · p. 8
5.2 Release Frame Extraction Result
- 94.6% accuracy and 95.0% mAP made RT-DETR the top release-frame detector; YOLO accuracy rose from 85.9% with YOLOv5 to 91.0% with YOLOv11.The authors selected RT-DETR for the proposed pipeline.Source: Table 3 · p. 9, §5.2 · p. 9
5.3 Results and Discussion
- CNN-based release-frame detectors had low classification accuracy and negative MCC, whereas only object-detection models yielded positive MCC.VGG-16, VGG-19, and ResNet-50 were reported as unsuitable for this task regardless of pose-estimation framework.Source: §5.3 · p. 9
- With OpenPose, RT-DETR raised accuracy from YOLOv8’s 72.73% to 75.00%, recall from 71.43% to 85.71%, and MCC from 0.4485 to 0.4781.OpenPose’s slower inference was noted as a constraint for real-time deployment.Source: §5.3 · p. 9
- 91.96% accuracy, 100% recall and TPR, and 0.8367 MCC were achieved by the MediaPipe–RT-DETR pipeline, the best overall combination reported.It also had 87.50% precision, 93.33% F1-score, and 20% FPR.Source: §5.3 · p. 10
- Misdetected release frames can propagate errors into elbow-angle estimation and produce incorrect action classifications.Figure 7 contrasts correctly classified deliveries, where both frames were detected accurately, with a misclassified example involving an inaccurate release frame.Source: §5.3 · p. 10
- Reliability depends on frame extraction and pose-estimation accuracy; validation across more diverse bowling styles and environments remains necessary.The authors identify this broader testing as needed to assess robustness under different match conditions.Source: §5.3 · p. 10
6 Conclusion
The system flags a delivery when the arm-angle change between shoulder and release frames exceeds 15°, and the authors report promising potential for real-time detection. They also identify limits in dataset diversity, reliability, and comparative evaluation.
6 Conclusion
- The human-labeled evaluation achieved a 100% true positive rate and 90% accuracy in distinguishing legal from illegal actions.Source: §6 · p. 10
- The system measures arm-angle change between shoulder and release frames and flags actions exceeding a predefined 15° threshold as illegal.Source: §6 · p. 10
- The dataset excludes female, side-arm, and left-handed bowlers; no direct computer-vision comparison is available, and 20% of illegal actions were classified as legal.Source: §6 · p. 11
- The authors say lightweight design and inference taking a few seconds support local or cloud deployment and potential live-match use, while future work targets broader data and fewer false negatives.Source: §6 · p. 11