Vehicle Detection in Aerial Traffic Monitoring

Abstract

This work describes a cascade detection of vehicles in Unmanned Aerial Vehicle (UAV) images and videos. There are some new approaches used in the detection. In particular, the Region of Interest (ROI) search is not only based on GIS and navigation data, but also employs visual method based on rapid image segmentation and road detection. The work also suggests doing ROI segmentation by the superpixel technique and trainable four-level cascade detector that uses artificial neural networks as classifiers. Characteristics of the being analyzed regions (combined superpixels) are based on geometric and texture features, as well as on deep features extracted from the image patches by nonlinear auto encoders. To improve the detection quality of the moving vehicles a separate stage of the detector based on optical flow analysis was introduced. Proposed detection algorithm was benchmarked on the real UAV videos and showed the sufficiently high accuracy. Performance of the algorithm allows supposing the on-board usage.

Automatic Object Detection, Labelling, and Localization by Camera’s Drone System

“…The construction system has two major components, which are the drone itself and a personal computer (PC). Table (1) contains the specifications of the modified drone prototype. The drone was built with a higher number of small systems.…”

“…Introduction Unmanned Aerial Vehicles (UAVs) have drawn a lot of interest for their image identification capabilities on both vehicles and humans. This interest has been prompted via military applications, such as aerial surveillance, as well as automatic target recognition in commercial and civil applications, including rescue missions and traffic search [1][2][3][4]. Various images or videos must be taken and analyzed in such applications, the majority of them must be processed in real-time.…”

“…The proposed system The proposed method involves the design of a drone for the real-time detection and labelling of objects from captured images/videos using a modified 1D-CNN. The block diagram in figure (1) describes the main steps of the proposed system. More details are explained in the following sections.…”