2020/03/09 by Shivam Prakash Gautam, Gautam, Shivam, Gregory P. Meyer +5
Computer Science · Engineering · #Video Surveillance and Tracking Methods #Autonomous Vehicle Technology and Safety #Traffic Prediction and Management Techniques
paper · pdf · doi:10.48550/arxiv.2003.04447
Accurate motion state estimation of Vulnerable Road Users (VRUs), is a\ncritical requirement for autonomous vehicles that navigate in urban\nenvironments. Due to their computational efficiency, many traditional autonomy\nsystems perform multi-object tracking using Kalman Filters which frequently\nrely on hand-engineered association. However, such methods fail to generalize\nto crowded scenes and multi-sensor modalities, often resulting in poor state\nestimates which cascade to inaccurate predictions. We present a practical and\nlightweight tracking system, SDVTracker, that uses a deep learned model for\nassociation and state estimation in conjunction with an Interacting Multiple\nModel (IMM) filter. The proposed tracking method is fast, robust and\ngeneralizes across multiple sensor modalities and different VRU classes. In\nthis paper, we detail a model that jointly optimizes both association and state\nestimation with a novel loss, an algorithm for determining ground-truth\nsupervision, and a training procedure. We show this system significantly\noutperforms hand-engineered methods on a real-world urban driving dataset while\nrunning in less than 2.5 ms on CPU for a scene with 100 actors, making it\nsuitable for self-driving applications where low latency and high accuracy is\ncritical.\n