Target Detection and Adaptive Tracking Based on Multisensor Data Fusion in a Smoke Environment

Abstract

Robust and accurate target detection and tracking in complex environments such as smoke are prerequisites for intelligent vehicle safe trajectory planning and navigation. Existing target detection and tracking methods have disadvantages, such as a high miss detection rate and apparent tracking deviation (position drift and size deviation of the tracking box) in a smoke environment. In response to these problems, we propose utilizing multisensor data fusion technology to improve the target detection recall and propose a new adaptive KCF tracking algorithm to improve tracking accuracy. We collected data from the environment of smoke to train the object detection network and designed a radar noise filter based on an empirical model to improve the accuracy and credibility of target detection. We then use a decision-level fusion strategy to fuse the image detection results and the radar filtering results. Finally, we utilize the fusion data to correct the tracking results online. The proposed method is evaluated on a self-developed intelligent vehicle platform, and the experimental results prove the algorithm's feasibility. This work provides a feasible solution to the problem of target detection and tracking in smoke environments.