Visual sensor networks (VSN) are networks of smart cameras capable of local image processing and data communication. Unlike traditional camera-based surveillance network in which cameras stream all image data to a centralized server for processing, cameras in VSNs form a distributed system, performing information extraction and collaborating on application-specific tasks. This thesis studies how complex vision tasks can be integrated with system resource constraints such as computation capacity, battery power and bandwidth, in two different VSN contexts. The first context is large-scale ad-hoc wireless smart cameras working on battery power which resembles the architecture of general wireless sensor networks. In this context we build geographic hash table based network protocols that are adapted to the nature of image sensors. These protocols decouple the event sensing from the camera location. Simulation results show that these protocols allow efficient distributed camera calibration and event-based constraint processing. The second context is smaller-scale static wired smart cameras with constant power supplies which can be found in public spaces that need surveillance such as airports and casinos. We study the performance advantages of applying probabilistic fusion methods in cross-camera object tracking. Object tracking based on a single feature type can produce high error rates due to environmental and view changes. We present a probabilistic object matching framework which employs multiple object features and a decision mechanism to combine results from multiple features. The framework builds matching probability distributions for each feature algorithm based on empirical data and combines these historical results into an aggregated result. Our experimental studies on realistic data show that while there is no single feature algorithm works best all the time, our probabilistic integration method on multiple features can almost always achieve better object matching accuracy than the best individual feature algorithm.