tshjustin / vehicle-image-recognition

Analyzing vehicles on roads using computer vision.

Traffic Images Vehicle Detection

Training of Model

• Traffic Images downloaded from API and annotated using Roboflow, which is then exported for training of model
• Import pre-trained model forom YoloV8 and fined-tuned (train) with exported data, then exported in ONNX Format
• Prepare ONNX model for inference

Tools	Description
Roboflow	Computer vision framework for data collection, annotation, preprocessing and augementation. Supports custom datasets and smart labeling with pre-trained model on COCO (Common Objects in Context).
YoloV8	Pre-trained model that comes with different weights and sparcity of model. Model can be imported and fined-tuned with custom dataset for specific use cases.
ONNX	Intermediary Machine Learning framework to convert between different ML frameworks to deploy accross differnet platforms. ONNX was used for easy deployment to Render cloud hosting & for comparison of model in different frameworks for a common platform.

Outcomes

Endpoint enacted that accepts a image URL. An example of the annoted image is shown below:

Confidence values for identified objects are also logged:

Class 0 : Score 0.906679630279541
Class 0 : Score 0.8510133028030396
Class 0 : Score 0.8255020380020142
Class 0 : Score 0.7501163482666016
Class 0 : Score 0.7322111129760742
Class 0 : Score 0.7169588804244995
Class 0 : Score 0.7063149213790894
Class 0 : Score 0.6881899833679199
Class 0 : Score 0.6813197135925293
Class 0 : Score 0.6607131361961365
Class 0 : Score 0.6455743312835693
Class 0 : Score 0.6336331367492676
Class 0 : Score 0.3735249638557434

Testing the ONNX model on test data yields an accuracy score of 0.706.

Images are hand annotated to ensure correct classfication of objects and control over classes.