With the ever increasing adoption of smartphones worldwide, researchers have found the perfect sensor platform to perform context-based research and to prepare for context-based services to be also deployed for the end-users. However, continuous context sensing imposes a considerable challenge in balancing the energy consumption of the sensors, the accuracy of the recognized context and its latency. After outlining the common characteristics of continuous sensing systems, we present a detailed overview of the state of the art, from sensors sub-systems to context inference algorithms. Then, we present the three main contribution of this thesis. The first approach we present is based on the use of local communications to exchange sensing information with neighboring devices. As proximity, location and environmental information can be obtained from nearby smartphones, we design a protocol for synchronizing the exchanges and fairly distribute the sensing tasks. We show both theoretically and experimentally the reduction in energy needed when the devices can collaborate. The second approach focuses on the way to schedule mobile sensors, optimizing for both the accuracy and energy needs. We formulate the optimal sensing problem as a decision problem and propose a two-tier framework for approximating its solution. The first tier is responsible for segmenting the sensor measurement time series, by fitting various models. The second tier takes care of estimating the optimal sampling, selecting the measurements that contributes the most to the model accuracy. We provide near-optimal heuristics for both tiers and evaluate their performances using environmental sensor data. In the third approach we propose an online algorithm that identifies repeated patterns in time series and produces a compressed symbolic stream. The first symbolic transformation is based on clustering with the raw sensor data. Whereas the next iterations encode repetitive sequences of symbols into new symbols. We define also a metric to evaluate the symbolization methods with regard to their capacity at preserving the systems' states. We also show that the output of symbols can be used directly for various data mining tasks, such as classification or forecasting, without impacting much the accuracy, but greatly reducing the complexity and running time. In addition, we also present an example of application, assessing the user's exposure to air pollutants, which demonstrates the many opportunities to enhance contextual information when fusing sensor data from different sources. On one side we gather fine grained air quality information from mobile sensor deployments and aggregate them with an interpolation model. And, on the other side, we continuously capture the user's context, including location, activity and surrounding air quality. We also present the various models used for fusing all these information in order to produce the exposure estimation.