Abstract:
Aerial object detection is becoming increasingly popular since images with a large field of view provide higher information per pixel than images with a short field of view. It is popularly used for pedestrian detection and crowd counting due to its importance in surveillance-related tasks. Traditional object detection algorithms such as convolutional neural networks alone, prove to be unsatisfactory at such heights since the objects such as pedestrians are blurry and too low resolution in the images to draw any reasonable inference. Pedestrian detection has significant difficulties in detection from high altitudes as humans are difficult to localize in a large perspective and feature identification for pedestrians is difficult since the object is localized in a very small portion of the image This report summarises work done in the first semester of Research track B.Tech project entitled Aerial pedestrian detection using vision based approach. Our work includes a literature review of 30 research papers on existing approaches for doing pedestrian detection from aerial view to find methods that can boost the performance of new models. We use the gathered information to develop auto-encoder based approach for image compression and also experiment with saliency based image segmentation to decrease the input size of input vector to neural network by keeping only the most relevant part of the image as input to a custom object detector trained on aerial images of pedestrians. We finally propose a vision based architecture for pedestrian detection and crowd counting, which will be implemented next semester.
