Multi Target Detection of UAV Video Based on Deep Convolution Neural Network

Abstract

Aiming at the existing multi-target detection algorithm in UAV shooting dynamic scene effect is not good, Color similarity problem, Athletes block each other. In this paper, a region compensation method based on fixed background is to propos detect multiple players in soccer video captured by UAV. Through Effective marking of targets in aerial images and the framework construction of deep convolution network. The difference processing of two adjacent frames, the changing and unchanging regions in the image are distinguished, and different regions are given different update rates and added to the background frame, in order to achieve faster background reconstruction. Experimental results in different video sequences, the detection accuracy of the algorithm is more than 90% on average. The algorithm processing has real-time, accuracy and stability of target detection.