An empirical performance analysis for on road object detection from Traffic videos: An image pre-processing perspective
Keywords:
YOLO, Mask-RCNN, Traffic Surveillance, Image pre-processing.
Abstract
This paper presents an empirical performance analysis of various object detection algorithms for the identification of on road objects from the perspective of different pre-processing techniques. The analysis is done on some real time traffic videos data captured from CCTV camera. The preprocessing techniques considered for analysis are background subtraction, denoising and smoothing methods. For background subtraction, two popular algorithms named Temporal Median Filtering and Canny Edge Detection are being utilized. The Temporal Median Filtering outputs the mask of the objects of interest by eliminating the background and Canny Edge Detection draws the outline of the objects of interest. Gaussian Blur technique is used for image smoothing and noise removal from traffic video. These transformations are applied on test videos of on road vehicle traffic collected manually and the results are obtained using state of the art methods. Apart from this analysis, a new trainingtesting model for quality object detection under different light conditions (day light, night and low vision) have also been proposed. The main idea here is to feed the detectors with images having less but meaningful features like object boundaries. This paper also presents a comparative analysis on state-of-the-art object detection algorithms such as YOLO and Mask-RCNN using transfer learning concept. Moreover, this paper also compares the performance of a segmentation model when fed with the data labelled with bounding boxes instead of the pixel level segmentation.
Published
2021-05-31
Section
Review Article
Copyright (c) 2021 International Journal of Advanced Engineering Research and Applications
![Creative Commons License](http://i.creativecommons.org/l/by-nc-nd/4.0/88x31.png)
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.