基于瓶颈神经网络的轨迹嵌入技术及其在飞行目标轨迹分类中的应用

Trajectory Embedding Based on Bottleneck Neural Network and Its Application on Trajectory Classification of Aircrafts

轨迹分类是利用目标运动轨迹识别出目标类型的技术。如何从轨迹数据中提取出可分性好的轨迹特征一直是轨迹分类的研究重点。本论文提出了基于神经网络的轨迹嵌入方法,从轨迹数据中提取可分性较好的轨迹特征。该方法首先从轨迹数据中提取局部和全局特征,组成高维度的特征向量;然后将这些高维度向量带入瓶颈神经网络(bottleneck NN, b-NN),得到低维度的超向量,称为t-vector。因为b-NN将高维度的特征向量投影到低维度的“通用坐标”空间中,对特征向量进行了校准和压缩,所以t-vector在超向量空间中具有较低的维度和较好的可分性。实验表明,t-vector能够提升分类模型5%以上的准确率,并使其检测代价值(detection cost function, DCF)较低,有效提高了飞行目标轨迹分类的性能。

Trajectory classification is a technique which classifies the objects based on their trajectory. The key of trajectory classification is to find the discriminative features that better dicriminate the class. This paper proposes a trajectory embedding method based on neural network to extract the discriminative features from trajectory data. This proposed method extracted the local and global features having high demension from raw trajectory data at first, and then obtained the low-demension supervector named t-vector using the bottleneck nerual network (b-NN). Due to the b-NN maps the high-dimension vectors into a low-deminson “common coordinate” space, the t-vector had low deminson and good discrimination when doing similarity computations in the supervector space. The experimental results shown the classification model with t-vector obtained more than 5% accuracy, and obtained lower detection cost function (DCF), which improved the performance of the trajectory classification.