Visual Place Recognition (VPR) remains a challenge in environments with commonly occuring repetitive features with no globally distinctive landmarks. This paper proposes an ensemble of Particle Filters (PF) for robust sequential VPR. The framework maintains a diverse ensemble of particle filters (PF) using complementary appearance and semantic observation models. From the real data we observed that a little deviation in position from the true position drasticallyreduces the similarity and such particles have unreliable variations, making it easier for particles to drift. We maintain replicas of the same observation models working at different temporal resolution (some sparingly using the observation model), which safeguards against an unreliable observation model drifting away the particles. We assume that different scenarios would have different observation models suited for the scenarios and hence we do not know which expertis best for a test scenario. We use particle confidence, inter PF consensus and trajectory smoothness as a metric to fuse the outputs of different PFs and geta reliable output against all types of scenarios. Using this reliable output, we inject a corrective probe into PFs whose particles have potentially drifted to bias the particles towards the correct positions, without affecting the PFs with correct predictions. The proposed framework is evaluated on three challenging benchmark datasets covering substantial viewpoint and environmental variations. Experimental results demonstrate improved localisation, robustness, trajectory consistency, and recovery from localization drift compared with conventional sequential VPR methods.