Home
Nowadays, the theme of autonomous cars is becoming increasingly important in the automotive industry, so this work consists of designing a miniature of it with the following three objectives:
A 1/10th scale electric car was used, with an on-board computer (Raspberry Pi 4), a neural network processor (Coral USB Accelerator), a camera, a distance sensor and a battery.
The road is tracked by an image processing which isolates the lines and determines the trajectory and curvature of the road.
The traffic sign detection algorithm uses a camera and three artificial neural networks to locate and classify signs.
- In the first network, a convolutional neural network compute additional information from the image such as contrast, edges, shape, colour, etc.
- Then with this inforamtions, a second network predicts the probable position of traffic sign in the image.
- The last network will then classify the type of sign at each predicted position
Obstacle detection is possible thanks to an ultrasound sensor on the front of the car.
Obstacle detection uses a simple process and therefore has a very low latency. However, objects of a certain shape and small size are not detected systematically.
Traffic sign detection is fast and reliable thanks to the robustness of the algorithm and the use of a processor specially designed for processing neural networks.
Road tracking is currently not fast and reliable enough, as image processing is very resource-intensive for the processor.
This work was done as part of my Bachhelor thesis. Find the final report here (only in french, sorry).
- Self Driving RC Car (by Zheng Wang)
- DeepPiCar (by David Tian)
- Curved Lane Detection (by kemfic)
First of all, we will build the car, the road and the signs.
Then we will install the raspberry.
Once everything is ready, the programming can begin.
I will start by explaining the structure of my program so that you can get off to a good start.
It will then be time to tackle the three main algorithms: