Two-level cascade model for tracking pedestrians using thermal infrared video information

Abstract

Thermal infrared pedestrian tracking is a challenging task due to factors such as energy attenuation, sensor noise, occlusion, and complex backgrounds. In this paper, we design a two-level cascade model that tracks pedestrians in a thermal infrared video by the coarse-to-fine strategy to improve the tracking accuracy and success rate. The base tracker in the first level of our model is initialized and fine-tuned to get the first representation of a target which is then used to locate the target roughly. Aiming at finely locating a target, the second level consists of modality-specific part correlation filters that can capture patterns of thermal infrared pedestrians. The outputs of part correlation filters are aggregated together by normalized joint confidence that can effectively suppress low confidence predictions to make a final decision. We adaptively update each part filter by a weighted learning rate and accurately estimate pedestrian scale by a scale filter to improve tracking performance. The experimental results on the PTB-TIR benchmark show that the proposed cascade tracker further emphasizes crucial thermal infrared features. Thus it can effectively relieve the problem of object occlusion. Our experimental results show the superiority of the proposed tracker over the state-of-the-art trackers, including SRDCF, GFS-DCF, MCFTS, HDT, HCF, MLSSNet, HSSNet, SiamFC_tir, SVM, and L1APG.