TASKS PERFORMED 🡆 Data Collection (Manual, Web Scrapping)✅, Exploratory Data Analysis✅, Data Augmentation✅, Model Training (ML Algorithms, CNN, Transfer Learning)✅, Hyperparameter Tuning✅, Model Validation✅, Deployment (UI Design, React and FastAPI)✅

App Demo

Upload images and make classification on:

• traffic_unrelated: Indicates an absence of traffic related situation
• congested_traffic: Indicates complete standstill caused by congestion or blockages or accidents
• heavy_traffic: Indicates high volume of vehicles causing slow moving traffic conditions
• moderate_traffic: Indicates moderate or steady flow of traffic with moderate number of vehicles
• light_traffic: Indicates minimal or very few vehicles on the road

For more details on the model training process, please visit the Kaggle project here.

• Prerequisites
• Demo
• Usage
• Model Summary
• Model Evaluation
• Acknowledgement

• You must have Python 3 installed on your system.
• You must have Git installed on your system.
• You must have Git LFS installed on your system (for cloning/downloading pre-trained model file).
• (Optional) You should have Node.js (npm) installed on your system (if using development server)

Step 1 - Upload Image

Step 2 - Crop Image

Step 3 - Result

1. Clone the repository (In terminal):

   git clone https://github.com/AbhashChamlingRai/traffic-image-classifier.git
   

2. Enter into the project directory:

   cd traffic-image-classifier
   

   Before following the below procedure, go to this link and download the model traffic_classifier.h5 and place it into the /traffic-image-classifier directory.

3. Install the required dependencies:

   pip install -r requirements.txt
   

4. Start Fast-API:

   Open /fastapi directory in your terminal. Then run the below command and wait until you see the message Application startup complete. in your terminal:

   uvicorn main:app --port 8000 --reload
   

5. Run React App:

   Open /react-app-build/index.html file in your browser. Upload images in the web app and make predictions.

6. (Optional) If you want to run development server:

   Open /react-app-development directory in your terminal. Then install project dependencies using npm install command. Wait until installation process completes, then run the development server by running the command npm start.

_________________________________________________________________
 Layer (type) Output Shape Param # 
=================================================================
 input_4 (InputLayer) [(None, 300, 300, 3)] 0 
 
 tf.cast_1 (TFOpLambda) (None, 300, 300, 3) 0 
 
 tf.math.truediv_1 (TFOpLamb (None, 300, 300, 3) 0 
 da) 
 
 tf.math.subtract_1 (TFOpLam (None, 300, 300, 3) 0 
 bda) 
 
 xception (Functional) (None, 2048) 20861480 
 
 batch_normalization_11 (Bat (None, 2048) 8192 
 chNormalization) 
 
 dense_3 (Dense) (None, 368) 754032 
 
 batch_normalization_12 (Bat (None, 368) 1472 
 chNormalization) 
 
 dense_4 (Dense) (None, 112) 41328 
 
 batch_normalization_13 (Bat (None, 112) 448 
 chNormalization) 
 
 dropout_1 (Dropout) (None, 112) 0 
 
 dense_5 (Dense) (None, 5) 565 
 
=================================================================
Total params: 21,667,517
Trainable params: 800,981
Non-trainable params: 20,866,536
_________________________________________________________________

The model needs input of a 300x300 pixel color image represented a NumPy array with RGB (3 channels). So, the input array must be of shape (1, 300, 300, 3). You don't have to worry about normalizing the pixel values (scaling by 1/255) because the first four layers of the model take care of it.

It generates prediction confidence scores represented by 5 consecutive numbers, each corresponding to a specific class: congested_traffic, heavy_traffic, light_traffic, moderate_traffic, and traffic_unrelated (in the specified order).

Data split distribution chart

The data was partitioned into two sets for training, and validation, each containing samples from five distinct classes. The split ratios are as follows: 85% for training, 15% for validation. Despite the limited data, the model successfully met my expectations:

For validation data:

score, acc = model.evaluate(validation_generator)
print('Test Loss =', score)
print('Test Accuracy =', acc)
Output:
8/8 [==============================] - 12s 1s/step - loss: 0.1525 - accuracy: 0.9468
Test Loss = 0.15245917439460754
Test Accuracy = 0.9468421339988708

Confusion Matrix:
[[350 15 1 11 1]
 [ 14 290 1 15 0]
 [ 1 1 345 11 2]
 [ 5 12 2 332 0]
 [ 4 0 5 0 482]]
Classification Report:
 precision recall f1-score support
 0 0.94 0.93 0.93 378
 1 0.91 0.91 0.91 320
 2 0.97 0.96 0.97 360
 3 0.90 0.95 0.92 351
 4 0.99 0.98 0.99 491
 accuracy 0.95 1900
 macro avg 0.94 0.94 0.94 1900
weighted avg 0.95 0.95 0.95 1900

For train data:

score, acc = model.evaluate(train_generator)
print('Test Loss =', score)
print('Test Accuracy =', acc)
Output:
43/43 [==============================] - 239s 6s/step - loss: 0.0883 - accuracy: 0.9695
Test Loss = 0.08829263597726822
Test Accuracy = 0.9695025682449341

Confusion Matrix:
[[2055 48 11 21 3]
 [ 57 1711 5 36 1]
 [ 4 3 2001 27 5]
 [ 18 35 14 1920 2]
 [ 3 1 4 4 2766]]
Classification Report:
 precision recall f1-score support
 0 0.96 0.96 0.96 2138
 1 0.95 0.95 0.95 1810
 2 0.98 0.98 0.98 2040
 3 0.96 0.97 0.96 1989
 4 1.00 1.00 1.00 2778
 accuracy 0.97 10755
 macro avg 0.97 0.97 0.97 10755
weighted avg 0.97 0.97 0.97 10755

Training and validation loss and accuracy graph

Special thanks to Asha Limbu for her work in creating the UI/UX design.

• Connect with Asha on LinkedIn or GitHub