Motion Prediction for First-Person Vision Multi-Object Tracking

Abstract

Tracking multiple independently moving objects with cameras mounted on moving robots is becoming increasingly common. However, most causal trackers rely on linear motion models that may be inaccurate in these scenarios. To overcome this problem, we present a real-time multi-object tracker based on the Early Association Probability Hypothesis Density Particle Filter with a prediction model that disentangles the motion of objects from that of the camera. Moreover, the prediction model allows us to intentionally reduce the video frame rate at which the tracker operates, with only a minor reduction in accuracy. Specifically, the model allows us to halve the processed frames while still outperforming alternative prediction models, including traditional linear motion predictors in moving-camera sequences. Experimental results show that the proposed model improves both accuracy and precision of the tracker.