Project Plan

General Plan

• Research
  • How to take output data from python script, send it through to the Raspberry Pi server, redact the output data, and have the data send to the app autonomously?
  • What are the best angles and camera settings to consistently and accurately detect the most amount of vehicles?
  • What sensors are most cost efficient and least intrusive installation wise for our use case? 
• Development 
  • Integrate a pre-trained vehicle detection model into the existing python code
  • Create an app using React Native for end users
  • Utilize redaction software for visual data privacy 

Task Breakdown 

• Project Management: Manages the overall project flow and handles external communications with stakeholders and advisors

  • Frank Genderson 

• Camera Systems: Researches and compares camera systems and the tools and architecture around them

  • Joris Wilson

  • Kavin Mohan

• Google Sheets Server: Creates the server to be run on Google Sheets

  • Frank Genderson

  • Kavin Mohan

  • Alan Manjarrez

• Vehicle Detection Software:  Creates original and leads in modifying existing python code

  • Joris Wilson

• React Native App Development: Programs and manages the user interface for the ROOST app

  • John Shea

  • Alan Manjarrez

• Meeting Minutes: Manages write-ups of meeting events 

  • John Shea

Concepts

Camera System

The camera system involves one camera and takes pictures over specified intervals.  These pictures are then sent through a redaction filter to increase privacy.   The photos then go through a pre-trained algorithm that detects if a car is in these spots or not.  This data is then sent through a server which is then sent to an app to display which spots are open.

Sensors 

We had considered using a multitude of different sensors including, ultrasonic, pressure, inductive, and thermal for car detection.  For the sensors, there would need to be one sensor per car detected, making it less cost-effective in the long run.  Inductive sensors would also require the inductive wire to sense each individual car.  

Vehicle Detection Software

We have looked between OpenCV and YOLOv5.  We tested both models and found out that that YOLOv5 was the better option because the model was pre-trained.  OpenCV could not detect as well for the images that we sent through the software.  However, OpenCV proved better to do the redaction tasks needed

Application Software 

We will be making an app using ReactNative and the Expo framework.  This will be using Javascript and Node.js using chocolatey.   Regardless, we needed the app to be developable for Android as that is easier to make.

Concept Selection

Design

 Analysis

Hardware Specifications

• Hikvision DS-2CD2955FWD-IS camera with 360° capture ability

• Raspberry Pi in order to act as a server for processing image data captured by the camera

Software Specifications 

• Acquire vehicle detection data from camera images.

• Data shall be parsed quickly and accurately.

• Code shall be efficient (to allow for quick data parsing, saving power, etc.).

• Visual Studio 

  • Programming integrated development environment (IDE).

  • Languages used: Python.

    • Python Libraries: OpenCV, os, torch.

• YOLOv5

  • Pre-trained vehicle detention model that can be intergraded into our existing python code.

• React Native 

  • JavaScript-based framework for developing mobile apps for Android and iOS.

• GitHub

  • Our GitHub page documenting our progress.

Test Plan

Out first round of testing was planned and accomplished in early October. The goal of this test was to take pictures of the vehicles parked in Babbio garage, and test them with the default YOLOv5 object detection software. This test was fairly simple, and somewhat unofficial, since the images would not reflect the quality and angles of the actually images we would be using with the actual camera. However, this test proved beneficial in testing if the detection software would work. This test had mixed results with the various images we took, but considering it was able to detect some of the vehicles, we considered this a small success.

Our second round of testing was planned in early November. For this test, the goal was to test the images received from the 360 degree camera, as well as the YOLOv5 software. The software used for the images was similar, the only difference was the new images reflected the actually location and angles we would have for the camera. This test proved to be more successful, as the two 180 degree views from the camera were able to capture a larger amount of cars in the parking lot from one camera. The results for this test were more successful than the previous test. Since we were able to adjust the parameters of the vehicle detection knowing what the quality and angle of the images we would be using, we were able to select the right parameters to detect the most amount of cars. The only issues we ran into for this test is because the images, which are at 180 degrees, are wide images, some of the vehicles at the edge of the image would not be detected since they were covered by other vehicles in the way. 

Going forward, we plan to take more images from the camera to test the images from different spots in garage, and run them through our image detection software. Moreover, we plan to add a feature where the license plates and people are blurred in order to protect the privacy of bystanders and personal information of the vehicles. In order to do this, we plan to use OpenCV's haar cascades, which are pre-trained models that are capable of detecting and blurring license plates and individual faces. Additionally, we plan to take the outputs from the vehicle detection software, and use it order to route the number of vehicles in the garage at any given time to a local server, and then to our app for the user.