Terrain aided navigation (TAN) is a promising approach to bound accumulated errors inherent to inertial navigation system by comparing terrain measurement with a reference map. Due to the non-linear nature of terrain, particle filters (PFs) are extensively studied for TAN because of its strong capability of dealing with non-linear problems. So far, most existing PFs for TAN manually select a fixed number of sampling particles during the entire filtering process. However, it can be highly inefficient, since the probability distribution of the state often varies drastically over time. To improve the efficiency, the Fox's adaptive PF based on Kullback-Leibler distance (KLD) is introduced for TAN, referred to as the normal KLD-PF here. In the normal KLD-PF, the number of sampling particles is adjusted online by KLD-sampling according to the size of state space. However, the normal KLD-PF has a fixed bin size, which easily causes the number of particles to surge at the early filtering stage. Thus, an improved KLD-PF with a variable bin size is proposed through limiting the total number of particles below certain value. Using a multi-beam sonar, simulation experiments with real underwater reference map demonstrate the efficiency of the proposed method.