Xing, Daitao (New York University), Tsoukalas, Athanasios (New York University Abu Dhabi), Evangeliou, Nikolaos (New York University Abu Dhabi), Giakoumidis, Nikolaos (New York University Abu Dhabi), Tzes, Anthony (New York University Abu Dhabi)

Siamese Adaptive Transformer Network for Real-Time Aerial Tracking

Scheduled for presentation during the Regular Session "Perception and Cognition" (ThA2), Thursday, June 23, 2022, 11:30−11:50, Bokar

2022 International Conference on Unmanned Aircraft Systems (ICUAS), June 21-24, 2022, Dubrovnik, Croatia

This information is tentative and subject to change. Compiled on April 24, 2024

Abstract

Recent visual object trackers provide strong discriminability towards accurate tracking under challenging scenarios while neglecting the inference efficiency. Those methods handle all inputs with identical computation and fail to reduce intrinsic computational redundancy, which constrains their deployment on Unmanned Aerial Vehicles (UAVs). In this work, we propose a dynamic tracker which selectively activates the individual model components and allocates computation resources on demand during the inference, which allows deep network inference on onboard-CPU at real-time speed. The tracking pipeline is divided into several stages, where each stage consists of a transformer-based encoder that generates a robust target representation by learning pixels interdependence. An adaptive network selection module controls the propagation routing path determining the optimal computational graph according to confidence-based criteria. We further propose a spatial adaptive attention network to avoid computational overhead in the transformer encoder, where the self-attention only aggregates the dependencies information among selected points. Our model achieves a harmonious proportion between accuracy and efficiency for dealing with varying scenarios, leading to notable advantages over static models with a fixed computational cost. Comprehensive experiments on aerial and prevalent tracking benchmarks achieve competitive results while operating at high speed, demonstrating its suitability on UAV-platforms which do not carry a dedicated GPU.