The increasing utilization of sensor devices in addition to human-given data make it possible to capture real world systems complexity through rich temporal descriptions. More precisely, the usage of a multitude of data sources types (devices, humans) allows to monitor an environment by describing the evolution in time of several of their dimensions through data streams. One core characteristic of such configurations is heterogeneity that appears at different levels of the data generation process: data sources, time models and data models. In such context, one challenging task for monitoring systems is to discover non-trivial temporal knowledge that is directly actionable and suitable for human interpretation. In this thesis, we firstly propose to use a Temporal Abstraction (TA) approach to express information given by heterogeneous raw data streams with a unified interval-based representation, called state streams. A state reports on a high level environment configuration that is of interest for an application domain. It is defined as a predicate involving data from one or several data sources. Such approach solves problems introduced by heterogeneity, provides a high level pattern vocabulary and also permits also to integrate expert(s) knowledge into the discovery process. Second, we introduced the Complex Temporal Dependencies (CTD) that is a quantitative interval-based pattern model. It is defined similarly to a conjunctive normal form and allows to express complex temporal relations between states. Contrary to the majority of existing pattern models, a CTD is evaluated with automatic statistical assessment of streams intersection avoiding the use of any significance user-given parameter. Third, we proposed CTD-Miner a first efficient CTD mining framework. CTD-Miner performs a incremental dependency construction. CTD-Miner benefits from pruning techniques based on a statistical correspondence relationship that aims to accelerate the exploration search space by reducing redundant information and to provide a more usable result set. Finally, as discovering pairwise significant time lag dependencies is a core operation in the CTD-Miner process, we proposed the Interval Time Lag Discovery (ITLD) algorithm. ITLD is based on a confidence variation heuristic that permits to reduce the complexity of the discovery process from quadratic to linear w.r.t a temporal constraint ∆ on time lags. To evaluate our approach, we implemented a motion simulation tool permitting to build data sets corresponding to a wide range of configurations. We also gathered data from a real world experiment using video cameras and real time video processing methods to build a real motion data set. Experiments showed that ITLD provides efficiently more accurate results in comparison with existing approaches. Hence, ITLD enhances significantly the accuracy, performances and scalability of CTD-Miner. The encouraging results given by CTD-Miner on our real world motion data set suggests that it is possible to integrate insights given by real time video processing approaches in a knowledge discovery process opening interesting perspectives for monitoring smart environments.