Introducing
Your new presentation assistant.
Refine, enhance, and tailor your content, source relevant images, and edit visuals quicker than ever before.
Trending searches
Electric & Autonomous Car
6
Design and implementation of a computer vision-based Autopilot for Autonomous cars
Self-driving cars are automobiles that consist of several sub-systems.
In this project, we are focusing on the design and implementation of a computer vision-based autopilot system.
Autonomous
project structure
computer vision :
Computer Vision
YOLOv4 or v4-tiny ?!
Object Detection
Semantic Segmentation
semantic segmentation
The following results were achieved using google Colab GPU: Tesla K80 with GPU memory of 12GB.
Results
Example (Different Environments):
U-net Results
Comparison Table:
Comparison
YOLO Results
Decision Making(Autopilot)
The Idea:
Modular
Genetic algorithm
-Wi ->weight of output i. -N->Total number of output length of data.
-acc ->accuracy. -f1->fitness function 1.
-f2 fitness function 2. -ref->desired output.
-ST-DIV ->The standard deviation of normal distribution.
Infertility Solution
Lane Tracer
the model
The Data
1 0.5820 0.7522 0.1816 0.2065
1 0.1092 0.7815 0.2287 0.3599
1 0.7940 0.7436 0.1085 0.1267
1 0.3232 0.7240 0.1736 0.1693
1 0.4525 0.7080 0.0877 0.1114
Data for the proposed GA
Lane Tracer
Length = 25 meters
Width = 0.6 meters
Total Captured Images = 11113 Image
End to End learning for self driving cars:
End to End
learning
NVidia
End to End
-Two models influenced by NVidia's End to End learning where developed .
our implementation
-Udacity and Carla simulation environments.
-Grand Theft Auto Vice City (2002) was the simulation environment.
simulation
creative solutions
1. Mutation calculation vectorization using mask binomial distribution, which resulted in a significant speed up.
2. Using just in time compiler (Numba JIT)to compile the class into machine code and then optimize the code resulted in a sharp reduction in the execution time.
3. Infertility detection and treatment protocol resulted in the avoidance of the deadends.
4. Tree method increased the diversity of the population which resulted in faster and better convergence.
5. Solving the problem of slow Learing in the early layers by introducing a slope in the mutation rate among the layers.
6. A transfer learning was done from a pre-trained YOLOv4 model on the COCO dataset using our input images. Then this data was used to train the YOLOv4-tiny model.
7. A small yet effective model, which is tolerant to overfitting for the lane tracer was developed.
Modular
1. Use of the state of the art Mish activation function along with normalization layers .
2. Get red of the analog nature of the Nvidia end to end model and replace it with a set of unique actions.
3. Creating a custom simulator form scratch .
4. Creating a custom memory system .
5. Introducing the ResNets and the VGG architectures to the Nvidia end to end model.
End to End
The following results were achieved using google Colab GPU: Tesla K80 with GPU memory of 12GB.
Results
The Proposed Genetic Algorithm
Lane Tracer
-ResNets-based model was found to be more accurate than VGG-based model.
-The VGG-based model was found to be faster than the ResNets-based model.
-The GPU delay time of the ResNets-based model is still relatively small and it may allow for a frame rate as high as 78fps.
End to End
Learning Curves
Milestones & Future Considerations
Conclusion & Future work
Conclusion
1. YOLOv4 tiny was found to be the most convenient object detection for our purposes given the current hardware limitations.
2. The modified U-Net model has shown higher accuracy compared to other models used for street semantic segmentation, but less than some models used in medical field.
3. Training phase of the genetic algorithm took more than 500 hours.
4. The small proposed model for lane tracing outperformed ResNets based models.
5. ResNets end to end model was found to be a better solution compared to the VGG end to end model.
6. Videos were captured to show the actual performance of the ResNets end to end and the small CNN model proposed for the lane tracer.
A blick to the Future!
1. More Capable GPUs & TPUs Processing Power.
2. More classes can be added to the computer vision models ( YOLO & U-Net).
3. Concatenate the U-Net model with ResNets End-to-End model.