Video Classification of Farming Activities with Motion-Adaptive Feature Sampling

Abstract

Recently, video has been applied in different industrial applications including autonomous driving vehicles. However, to develop autonomous farming vehicles, the video analysis must be targeted for specific farming activities. So an important first step is to classify the videos into their specific farming activity. In this paper, we propose a video classification framework that includes two branches that process videos differently based on their motions. A gradient-based method is proposed for separating videos into two subsets which are then processed by different feature sampling strategies. The result shows that two motion-based feature sampling strategies provide more efficient features; thus better classification performances are achieved. We also discuss how the feature sampling strategy influences the classification accuracy and the computational efficiency. In addition to farming videos, this proposed system can also be applied to classify videos captured from various camera movements, such as hand-held or first-person cameras.