Behavioural simulation of agents representing humanoid characters has spread to many areas in recent years. A part of such simulations are crowd simulations, where large numbers of agents move and interact at the same time. Finding a suitable level of individual agent complexity so that large simulations are possible and suitable behaviour is reached, is challenging. In addition, executing such a simulation in real-time is problematic. In my work I developed a real-time application in Unity game engine which makes use of a number of main techniques and approaches for heterogeneous crowd simulations, such as modular architecture, environment sensing, obstacle avoidance, finite state machines for behaviour modeling, animator for animation visualisation etc. I thoroughly described and presented those approaches and techniques and commented on the results obtained in several different scenes which represent specific real-world situations.