1 / 8

INNOSYS Award Post

INNOSYS: 2nd Best Project among 34 final year projects at Rajarata University

Overview

Real-Time Traffic Sign Recognition is an on-device Android system (SignMe) for detecting and announcing Sri Lankan traffic signs in difficult conditions such as night, rain, and partial occlusion. The pipeline covers condition-aware preprocessing, YOLOv11n-seg localization (mAP@0.5 ≈ 0.85), a lightweight CNN for 20 sign classes, and SimpleNLG voice alerts with duplicate suppression. We trained on a custom Colombo road dataset annotated in Roboflow. The project placed 2nd among 34 final-year projects and was published on ResearchGate.

Role
Research lead / corresponding author
Team
Team InnoSys (5), Rajarata University of Sri Lanka
Timeline
2024 · Published research
My contribution
Led the research pipeline: adaptive preprocessing, YOLOv11 segmentation, CNN recognition, SimpleNLG voice alerts, TensorFlow Lite Android deployment, and the IEEE conference paper.

Problem

Most TSR systems are tuned for daylight and rely on visual overlays that compete for attention. Sri Lankan roads, lighting, and signage are also underrepresented in public datasets like GTSRB, so off-the-shelf models underperform locally.

  • Drivers miss or misread signs under distraction, low light, rain, glare, and awkward angles
  • Classical color/shape segmentation collapses when illumination shifts
  • Visual-only alerts add cognitive load instead of reducing it
  • Commodity phones must run inference without a cloud dependency

Pipeline

All four research stages are open below with their result videos. Each clip plays when you scroll to it. Use the custom controls for play/pause, seek, mute, and fullscreen.

  • Four research modules: adaptive preprocessing → YOLOv11 segmentation → CNN recognition → smart auditory alerts
  • Full pipeline runs on-device in the SignMe Android app via TensorFlow Lite + OpenCV
1

Adaptive Preprocessing

PSNR > 30 dB in 8/9 cases

Classify lighting and degradation, then enhance only the frames that need it.

Before detection, each camera frame is classified by lighting (day/night) and degradation (dark/blur). Targeted brightness and contrast fixes are applied only where they help, so night and glare scenes stay usable without wasting compute on clean frames.

  • Environment detection: day vs night
  • Degradation detection: dark appearance and blur
  • Selective enhancement focused on regions that matter for recognition
  • Evaluated on 9 low-light cases (mean PSNR ≈ 32.95 dB; hardest case 27.81 dB fell below 30 dB)
2

YOLOv11 Segmentation

mAP@0.5 ≈ 0.85

Localize traffic signs in each frame and return crop boxes for recognition.

A fine-tuned YOLOv11n-seg model localizes traffic signs in the enhanced frame and returns bounding boxes for cropping. The nano variant keeps inference mobile-friendly while deep segmentation handles lighting and occlusion better than classical color/shape methods.

  • YOLOv11n-seg via TensorFlow Lite with OpenCV/Java on Android
  • Outputs (x1, y1, x2, y2) logged to MySQL for history and recognition
  • Training: 50 epochs, batch 16, 80/20 split, Adam with dynamic LR
  • Precision ≈ 0.84; recall peaks near 0.95; mAP@0.5:0.95 rises to just above 0.68
3

CNN Classification

~100% accuracy by epoch 10

Classify each crop into one of 20 Sri Lankan traffic-sign categories.

Cropped sign regions are resized to 30×30, normalized, and classified by a sequential CNN into 20 classes. The model is small enough for phones yet converges quickly with strong train/validation agreement and minimal overfitting.

  • Architecture: Conv 32×2 → pool/dropout → Conv 64×2 → dense 256 → softmax 20
  • Adam optimizer + categorical cross-entropy; 15 epochs, batch 32
  • 80/20 train–test split via scikit-learn
  • Train and validation accuracy both near 100% by ~epoch 10
4

Smart Auditory Alerts

Priority + no repeats

Announce high-priority signs once, and skip duplicates.

Recognized signs feed a SimpleNLG module that builds short, grammatical voice alerts. A history list blocks duplicates, and critical signs (e.g. Stop, Pedestrian Crossing) jump the queue so the phone informs without flooding the driver.

  • Rule-based SimpleNLG sentences delivered through Android TTS
  • Sign-history comparison skips redundant announcements
  • Priority ranking for safety-critical categories
  • Road tests confirmed clear alerts with low detection-to-announcement latency

Sign Classes

The recognition CNN covers twenty Sri Lankan traffic-sign classes grouped as warnings, speed limits, prohibitions, and a special pedestrian sign.

Dataset

Public datasets miss local context, so we built a custom onboard-camera corpus around Colombo (day, day-with-dark, night, and night-with-light) and annotated it in Roboflow for both segmentation and recognition.

  • 4,000+ annotated road-scene frames managed in Roboflow
  • 20 traffic-sign categories across DWS, SLS, PHS, and OSD groups
  • 80/20 train–validation for YOLOv11; 80/20 train–test for the CNN
  • Augmentation via flip, scale, and color jitter for lighting robustness

Results

Evaluation covered preprocessing quality, segmentation precision/recall, recognition accuracy, and real-driving usability of the voice alerts.

  • Preprocessing: PSNR > 30 dB in 8 of 9 low-light tests (mean ≈ 32.95 dB)
  • Segmentation: mAP@0.5 ≈ 0.85; precision ≈ 0.84; recall peaks near 0.95
  • Stricter mAP@0.5:0.95 climbs from ~0.35 to just above 0.68
  • Recognition CNN: train and validation accuracy converge near 100% by epoch 10
  • Auditory module: low detection-to-announcement latency with effective duplicate skipping in road tests

Recognition

The work placed 2nd at the university research competition and was published as an IEEE conference paper, Real-Time Traffic Sign Recognition Using Computer Vision Techniques, with the full text on ResearchGate.

  • 2nd Place among 34+ final-year projects (INNOSYS)
  • IEEE conference paper with ResearchGate publication ID 410938287
  • Participated in IEEE Innovation Nation Sri Lanka 2024 (North Central Provincial Competition)
  • Corresponding author: I. D. I. L. Senavirathna, Department of Computing, Rajarata University of Sri Lanka
  • Live expo demos of SignMe with the full team and supervisor

Stack

Python for model training; Java/Android Studio + TensorFlow Lite + OpenCV for on-device inference; Roboflow for annotation; SimpleNLG for voice phrasing; MySQL for session history.

YOLOv11CNNOpenCVTensorFlow LiteAndroidJavaPythonRoboflowSimpleNLG

More projects

Explore other work in the portfolio

View all