Efficient Moving Target Detection Using Resource-Constrained Neural Networks

Abstract

In recent years, the widespread use of autonomous vehicles, such as aerial and automotive, has enhanced our abilities to perform target tracking, dispensing our over-reliance on visual features. With the development of computer vision and deep learning techniques, vision-based classification and recognition have recently received special attention in the scientific community. Moreover, recent advances in the field of neural networks with quantized weights and activations down to single bit precision have allowed the development of models that can be deployed in resource-constrained settings, where a trade-off between task performance and efficiency is accepted. In this work we design an efficient single stage object detector based on CenterNet containing a combination of full precision and binary layers. Our model is easy to train and achieves comparable results with a full precision network trained from scratch while requiring an order of magnitude less FLOP. This opens the possibility of deploying an object detector in applications where time is of the essence and a graphical processing unit (GPU) is absent. We train our model and evaluate its performance by comparing with state-of-the-art techniques, obtaining higher accurate results and provide an insight into the design process of resource constrained neural networks involving trade-offs.