tizo

github accounts: @titusnicolae @zoliszeredi

project page: https://github.com/HackTM2016/tizo

category: SmartCity table No: 31

Inspiration

there are a lot of new traffic lights and cameras installed in Timisoara as far as we can tell the cameras won't have computer vision to analyse the video

What it does

detects the number of cars in a video stream and controls a traffic light

How we built it

used a deep learning framework http://pjreddie.com/darknet/ installed it on NVIDIA Jetson TK1 embedded platform altered the video file to simulate different amounts of car traffic (right now trying to light up some LEDs for the traffic light)

Challenges we ran into

installing the OS, cuda and openCV on NVIDIA TK1

Accomplishments that we're proud of

we got the network to work and draw bounding boxes around the cars

What we learned

the training data set was mostly street level images videos of traffic on youtube is usually shot from higher up the network doesn't recognize the cars from such a different angle from the training set also if the video is blurry and the cars are small the network doesn't detect them, it was trained with images shot closer up to the cars

What's next for tizo

improve the training set so it's closer to the footage that the camera will see port to Movidius Fathom + raspberry pi try to install it in the city somewhere to benchmark the accuracy

Built With: C, CUDA, OpenCV, machine learning, deep learning, computer vision, NVIDIA Jetson TK1, webcam

Built With

• c
• cuda
• deep-learning
• machine-learning
• python
• yolo