Performance of OpenCV HOG Model for Various Image Quality Changes

Abstract

Pedestrian detection is a crucial task for computer vision specifically in the application of self-driving cars and road safety. Accurate and robust pedestrian detection models are essential for ensuring the safety and efficiency of a range of applications such as autonomous vehicles and surveillance systems. I use a data set from Kaggle to test a Histogram Oriented Gradients (HOG) model on different image quality. Using OpenCV I alter image characteristics such as contrast, brightness, and blur. Then I test these images on the HOG model and calculate the accuracy for images with and without pedestrians. After the changes to the images, the model performs significantly worse, especially for the blurry images and the images with low brightness. However, there are some cases in which changing the image characteristic does improve the model from the baseline. For example, increasing the contrast helps the model correctly identify the presence of a pedestrian. This is a good starting point for detecting pedestrians; however further iterations are needed before using this on the road.