Background
A geofence is a small geo-graphic area that is defined to generate a location event as soon as a user enters or leaves this geofence.
Geofence applications are used extensively in advertisement industry to target customers based on their location. It has many use cases in transportation industry as well. For example, issuing location-based notifications to drivers, geo-encoding areas, fetching traffic rules information etc. A good solution could help replace the need of high-end camera-based solutions for ADAS systems.
Problem
There are basically three methods to detect whether a point is in geofence or not. They are Ray casting, TWC (Triangle Weight Characterization) and winding number method.
Ray casting method detects whether a point is inside or outside by projecting an infinite ray from the point and count the number of intersections of each side. TWC method divides the given shape into triangle and determines iteratively whether the given point is inside the shape or not. Winding number method counts the number of times the polygon winds around the point. The point is outside when winding number is 0, otherwise it is inside.
Limitations identified are as follows:
- The computational complexity of these algorithms increases in proportion to the number of sides of the polygon.
- Geofencing algorithms like Ray casting methods being running continuously, there is a chance of increased battery drain and processing power of devices which are mainly battery powered.
- If the geofencing algorithms is implemented in the server, then that could result in increased network traffic, and it limits the use cases to only areas with good network connections.
Solution
There has been much research into solving these problems using machine learning algorithms. Google has patented a solution based on decision tree. There is also a pre-work done based on neural networks for geo-fence. In this thesis we investigate which machine learning algorithm is best suited for this particular use case and study the impact on performance on a system. With this, we will be able to reduce computational complexity on geo-fences with 4 or more sides, reduce impact on the system, provide geo-fences with little or no internet access.
Goal of the thesis
The goal of the thesis is to conduct a study on which machine learning algorithm along with their hyper parameters could provide the best results and propose an architecture which could overcome the current limitations in the system.
Desirable expertise
- C++
- Python
- Tensorflow
- Linux
- Basic Android development could be helpful in prototyping the solution
Kick-off date
As soon as possible
Additional info
The scope can be flexible and adapted to 1-3 students.
Some reference links:
