Pattern recognition methods in case of visual information semantic integrity violations

We consider computer vision issues with a view to embed it into vehicles and automate the traffic process by detecting of objects and road users. Three basic stages of pattern recognition in the image are enumerated: image preparation for analysis, analysis, object classification. The search of methods that improve pattern recognition quality and provide visual information semantic integrity was carried out respectively for each stage. We propose to use normalization for the first stage, which allows making image objects insensitive to the light changes. For the second stage clustering method based on particle swarm optimization and k-means was developed that provides automatic tuning of clustering parameters. During the third stage Haar cascades with normalized training samples are proposed for application providing object unification and giving an opportunity to use less amount of training samples. Car images provided by Stanford University laboratory and publicly available were used for training and testing. To assess the effectiveness of the developed pattern recognition algorithm, 300 test images were blurred. We compared the results of the proposed algorithm operation with Haar cascades operating results without normalized training samples and without preparation for classification. Haar cascades with non-normalized training images provided correct car recognition in 8% of cases, meanwhile, the proposed algorithm recognized 72% of cases including those 8% of images. Visual information semantic integrity preservation is an important aspect in context of road traffic, because incorrect object detection can cause fatal consequences. The proposed algorithm of image analysis reduces the probability of error occurrences during pattern recognition.