Vinsamlegast notið þetta auðkenni þegar þið vitnið til verksins eða tengið í það: http://hdl.handle.net/1946/39345
To help combat global warming by encouraging eco-friendly culture and behavior, a computer vision system was implemented to detect vehicles, read their license plates, and keep track of the number of passengers within. This thesis focused on designing, implementing, and evaluating functionality that made it possible to optimize the system with the hope of decreasing processing time and making the system more suitable for use on edge computers. The main goal was to decrease processing time by either not trying to read license plates that are unlikely to be read correctly or by skipping a few frames of the video from being evaluated by the system at all, given a certain condition. The main variables used to determine if a license plate should be attempted to be read were its distance from the camera, its angle to the camera, its count of identical readings, and whether the vehicle it belonged to had been marked as a stationary vehicle. The variables used to decide if frames should be skipped were whether the license plate of any vehicle in the current frame was within a certain distance, or within a certain count of repeated readings. Experiments were performed using the YOLOv4 object detection algorithm for all of those optimization variables along with the combination of the distance and count variables, which were used both for deciding if to read a license plate or to skip frames. Experiments were also performed using the Faster R-CNN object detection algorithm but only with the distance variable to decide if a license plate should be read.
The results from the experiments indicate that using the Faster R-CNN object detection algorithm is too slow for the desired purpose. Regardless of the detection algorithm used, using distance and license plate count variables to decide whether to read a plate only decreased the processing time minimally. Using a count of identical readings and stationary vehicle marking to decide whether to read a plate, also only decreased the processing time minimally except for videos where stationary vehicles were being detected in the background. In those cases they both resulted in great decreases in processing time but at the cost of accuracy in the amount of license plates found throughout the video, which can be explained by the amount of repeated readings of the stationary vehicles. Combining the distance and plate count values when deciding whether to read a plate also only decreased the processing time minimally except when stationary vehicles were present, then they resulted in a great decrease in processing time with lowered accuracy in the amount of license plates found throughout the video.
Skipping a few frames of the video from being evaluated by the system returned much better results. Frame skip using distance values resulted in both a substantial decrease in processing time with a slight decrease in accuracy, and great accuracy with a slightly less substantial decrease in processing time. Frame skip using plate count values gave a similarly substantial decrease in processing time as using distance values did, with slightly better accuracy. Finally, combining the two resulted in the most decrease in processing time, of around 50%, while still remaining at an acceptable accuracy in the amount of license plates found throughout the video, or around 75%.
|Go-Green or Go-Home.pdf||13.23 MB||Opinn||Heildartexti||Skoða/Opna|