Mining temporal sequence patterns using association rule mining algorithms for prediction of human activity from surveillance videos

Abstract

One of the most interesting and important open issues in the automated video surveillance community is the analysis of human activities which requires a higher level of understanding. A framework was introduced for the prediction of human activity using temporal sequence patterns. In this framework, Sequential Pattern Mining (SPM) converted the complex symbolic sequence in the video into frequent itemsets using the Apriori algorithm. Apriori algorithm has to scan the video multiple times to find the frequent itemset that results in high execution time and memory consumption problems. To solve these problems, various association mining algorithms such as Rapid Association Rule Mining (RARM), Equivalence class clustering, and bottom-up lattice traversal (Eclat), diffset Eclat (dEclat) and Frequent Pattern-growth (FP-growth) algorithms are introduced for mining frequent itemsets. RARM speeds up the mining process at the minimum support threshold by using the Support-Ordered Trie Itemset (SOTrieIT) tree structure. The performance of frequent itemset mining is enhanced by using Eclat where depth-first approach is accommodated in it to find the frequent itemsets. However, the pruning technique in Eclat is more difficult. So, a dEclat algorithm is used where a depth-first search is performed to find frequent itemsets. But, dEclat is not more suitable for the sparse database. So, an FP-growth algorithm is introduced where a compressed data structure called FP-tree is used that solves the time and memory problem of the Apriori algorithm. The mined frequent itemsets are normal human behavior and the remaining activities are abnormal human activity. The experiments are carried out to prove the effectiveness of the RARM, Eclat, dEclat, and FP-growth for mining frequent activities in the video.