Milestone 4
Beta Demonstration of Optimized Design
Beta Demonstration of Optimized Design
Subsystem 1 Demonstration and Optimization
Subsystem 1 Demonstration and Optimization
The above is a video demonstration of our final prototype demonstrating our system tracking a drone. This video is showing the image detection from the perspective of the system (this is the video with the big blue line in the middle). We are using a webcam to record video, where we then use image detection to identify if there is a drone on screen. Once we believe if there is one we process where it is on the screen, and then control the motor to either make the system rotate clockwise or counter clockwise based on the position of the drone. The main optimizations that occurred throughout the prototyping process was the effectiveness of our image detection. We were able to improve our image detection's effectiveness by training the image detection model with more images to make the model more accurate. In addition, to optimize the model for the Raspberry Pi, we trained a quantized MobileNet model and converted it to TensorFlow Lite. We also properly calibrated the motor to not move to fast, as it would have to over correct if it passed the drone, while also not being too slow to not be able to follow the drone.
The above is a video demonstration of our final prototype demonstrating our system tracking a drone. This video is showing the image detection from the perspective of the system (this is the video with the big blue line in the middle). We are using a webcam to record video, where we then use image detection to identify if there is a drone on screen. Once we believe if there is one we process where it is on the screen, and then control the motor to either make the system rotate clockwise or counter clockwise based on the position of the drone. The main optimizations that occurred throughout the prototyping process was the effectiveness of our image detection. We were able to improve our image detection's effectiveness by training the image detection model with more images to make the model more accurate. In addition, to optimize the model for the Raspberry Pi, we trained a quantized MobileNet model and converted it to TensorFlow Lite. We also properly calibrated the motor to not move to fast, as it would have to over correct if it passed the drone, while also not being too slow to not be able to follow the drone.
Subsystem 2 Optimization
Subsystem 2 Optimization
Subsystem 2, our subsystem that is in control of the jamming and RF detection, has been going through steady changes throughout the life of our project. Most of the changes since the alpha stage were incremental changes that increased the robustness of the detection over time. This includes transitioning the code from a Jupyter Notebook to a Python scrpt and adding parameters to adjust the detection thresholds more easily. These improvements were able to reduce the amount of false positives during detection and making the detection more accurate in guessing a center frequency. Additionally, the script will also alternate between two different local oscillator frequencies, allowing for the sdr to search a wider part of the spectrum for the presences of the drone.
Subsystem 2, our subsystem that is in control of the jamming and RF detection, has been going through steady changes throughout the life of our project. Most of the changes since the alpha stage were incremental changes that increased the robustness of the detection over time. This includes transitioning the code from a Jupyter Notebook to a Python scrpt and adding parameters to adjust the detection thresholds more easily. These improvements were able to reduce the amount of false positives during detection and making the detection more accurate in guessing a center frequency. Additionally, the script will also alternate between two different local oscillator frequencies, allowing for the sdr to search a wider part of the spectrum for the presences of the drone.
System level Optimization and Integration
System level Optimization and Integration
Below you can see an earlier design of our antenna position, which we later changed to angle the antenna up so it would be more in line with the camera angle. In addition to the antenna offset, we also moved power input from multiple sources, into one power distribution block which can be seen mounted to the underside of the device. This allows for more control of power going into the device. Lastly, we also were having problems with our initial motor, as it was having problems moving the entire system, so we had to get a new motor which now works much better and much more consistently.
Below you can see an earlier design of our antenna position, which we later changed to angle the antenna up so it would be more in line with the camera angle. In addition to the antenna offset, we also moved power input from multiple sources, into one power distribution block which can be seen mounted to the underside of the device. This allows for more control of power going into the device. Lastly, we also were having problems with our initial motor, as it was having problems moving the entire system, so we had to get a new motor which now works much better and much more consistently.
November, 2020
April, 2021