Direct Determination of Object Trajectories: An Alternative to Traditional Object Detection and Tracking in Autonomous Vehicles

Abstract

Proposed and evaluated is the viability of a novel approach to the problem of 2- dimensional object trajectory calculation within the context of autonomous vehicle navigation. Numerous convolutional neural network architectures have been developed for this application, but no matter the architecture, nearly all have been designed to operate through discrete consideration of each image. The task of locating objects is always performed anew for each input. Object movement is then determined through post-processing.1 It is argued that this is both unnecessarily computationally complex as well as neglectful of useful time-series data within the incoming images. For autonomous driving, full-featured object tracking is used primarily in collision avoidance. Performing this task requires only the trajectories and sizes of objects, and thus, current neural networks could be simplified by focusing solely upon these features. Treating object tracking as a perturbation problem, thus making explicit use of the time-series information available to the vehicle, is proposed to accomplish this task. In particular, utilizing differential inputs to a simplified network is considered. To test this perturbation-based neural network’s efficacy, it is compared in training and testing to pre-existing architectures also used for 2D object tracking, Deep Convolutional Neural Networks using post-calculation and LSTM-enabled Recurrent Neural Networks. All models are trained on the same set of 2-dimensional annotated videos to ensure uniformity. Throughout this process, measures of accuracy, network complexity, and training capacity are collected for both the perturbation-based network and existing network architectures. These metrics facilitate assessment of the relative merits and drawbacks of this novel perturbation approach. When comparing the perturbation-based model to existing neural network implementations, one generally sees both a slight reduction in requisite architecture complexity as well as an increase in training and testing aptitude. Consequently, this methodology provides an ideal alternative to contemporary neural network approaches when performing object detection for collision avoidance, particularly when processing power is limited.