A Single-Shot Approach Using an LSTM for Moving Object Path Prediction

Abstract

This work presents an analysis of predicting the future path of moving objects from a moving camera on traffic scenes with an LSTM architecture in a single-shot manner. Path prediction allows us to estimate the future locations of an object in a given space and is useful in important applications such as surveillance, abnormal behaviour detection, crowd behaviour analysis, traffic control and currently in driver assistance (ADAS) or collision avoidance systems. Normal approaches use the last tobs positions of an object observed in video frames to predict its future path as a sequence of position values. This can then be treated as a time series. LSTM architectures are known for reaching good performance when dealing with time series. We evaluate path prediction across three types of objects (pedestrians, vehicles and cyclists), four prediction horizons (5, 10, 15 and 20 frames ahead) and two different perspectives (image coordinate and birds-eye view). The approach described in this work reached an Average Displacement Error (ADE) of 0.01m for pedestrians, 0.06m for vehicles and 0.02m for cyclists and an average Final Displacement Error (FDE) of between 0.016m and 0.15m for near-future prediction using an LSTM architecure with relative tracklet positioning.