The main aim of the thesis is to explore computer vision based solutions to the reduction of false alarms in surveillance networks. More specifically, the problem of false alarms triggered by spiders, which contributes to a substantial percentage of nuisance alarms, is addressed. In an automated surveillance setup in which motion events trigger alarms, the percentage of false alarms raised by spiders can range from 20 50% depending on the season of the year, lighting conditions, camera type and other environmental factors. These alarms not only (a) increase the workload of human operators validating the alarms but also (b) increase labor costs associated with regular cleaning of the lens to avoid frequent build up of spiders/cobwebs. In this thesis, a novel and an economical method to reduce the false alarms caused by spiders is proposed by building a spider classifier intended to be part of the video processing pipeline for intruder detection systems. The proposed method, which uses a feature descriptor obtained by early fusion of image blur and texture, is suitable for real-time processing and yet comparable in performance to more computationally costly approaches like SIFT/RootSIFT with bag of visual words aggregation. The performance of the binary classifiers developed based on several visual features is comprehensively investigated. The proposed method can eliminate 98.5% of false alarms caused by spiders with a false positive rate of less than 1%, thereby reducing the workload of the surveillance personnel validating the alarms. This also optimises the usage of police resources, especially in situations where the event triggered due to the spider is not dismissed by an operator in time, resulting in police notification. The classifier confidence score also provides cues for prioritising events to be addressed and could be further used to actuate a mechanical wiper which might be used in clearing the spider webs remotely.