Optimization Design of Multi-object Tracking Algorithm Based on Roadside Cameras

Abstract:

In response to the limitations of current multi-object tracking algorithms in handling roadside traffic scenarios， a multi-object tracking algorithm based on roadside cameras was proposed in this paper. First， the one-shot tracking algorithm chosen and a neural network based on FairMOT was built to simultaneously generate both object detection results and appearance feature results， thereby enhancing the real-time performance of the algorithm. Then， a novel data association method was adopted to lessen the effect of occlusion on the tracker. After that， a new motion similarity measurement called buffered intersection over union was introduced to compensate for the errors caused by linear motion prediction models. Subsequently， a velocity-based discriminative algorithm for removing lost trajectories and a history-based position matching algorithm to retrieve the identities of occluded trajectories over lengthy periods of time were developed. Experiments were conducted on the UA-DETRAC public multi-object tracking dataset to verify the effectiveness of the algorithm. Additionally， to demonstrate the applicability of our algorithm in real-world roadside environments， real roadside scene data were collected on open field in the National Intelligent Connected Vehicle （Shanghai） Pilot Demonstration Zone. Finally， comparative experiments between the algorithm proposed and SORT， DeepSORT， ByteTrack and FairMOT algorithms were conducted using real-world roadside scene data. The experimental findings indicate that the proposed algorithm performs better than other algorithms in terms of identification F-score， ID switch， fragmentation， mostly tracked， mostly lost， and multiple object tracking accuracy.