Fusion of Optical Flow and Histogram of Oriented Gradient Features in Abnormal Activity Recognition Mahasweta Joshi 1 and Jitendra Chaudhari 2 1 Birla Vishvkarma Mahavidhyalaya, V. V. Nagar, Gujarat, India Email: mjjoshi@bvmengineering.ac.in 2 CHARUSAT University, Changa, Gujarat, India Email: jitendrachaudhari.ec@charusat.ac.in Abstract—Nowadays abnormal action and event recognition is very encouraging research topic in surveillance applications. There are three main component of abnormal activity recognition. The components are: (i) Feature representation approaches (ii) pattern recognition models and (iii) performance evaluation strategies. Compare to second and third component, first component: Feature representation is very important. It should have robust appearance and motion information. This information of the features plays very important role in video analysis. In this paper, fusion of features has been performed to find more accurate recognition of abnormal activity. Optical Flow and HOG (Histogram of Oriented Gradients) features has been fused to get more accurate feature vector. Upon this vector rule based recognition method has been applied for classifying abnormal event. This proposed algorithm has been implemented on UCF, BEHAVE and UMN dataset. The result shows that with fusion of features more accuracy has been achieved in finding starting frame of abnormal activity in videos. Compare to existing algorithm, proposed algorithm achieved 16%, 1.8% and 21% more accuracy in UMN, BEHAVE and UCF dataset respectively. Index Terms— Optical flow, Histogram of Oriented Gradients, UCF dataset. I. INTRODUCTION Video surveillance is promising research area in current world for asset and personnel safety. But we need automation in it. It is so because this can reduce the manual workload. Object detection, tracking, person detection, event recognition and activity recognition are different areas where current researchers are focusing. Public area can be monitored with the help of this applications and with less human power. However, detecting an abnormal activity from a video is a challenging problem that is extremely contextual [1]. Abnormal activities are an anomalies of the object behaviour from the usual behaviour. This includes object in uncommon place, uncommon pattern of motion such as movement in incorrect direction, illegal turns in traffic, object access or entry in private zone, fighting among the persons, left bag, unexpected movements, let fallen object or any type of uncommon occurrence. For one event it may be a normal in one scenario and abnormal in another scenario [2]. The activity recognition system is primarily having following steps:  Optional step: Pre processing  Important step: Extraction of features Grenze ID: 01.GIJET.8.2.508 © Grenze Scientific Society, 2022 Grenze International Journal of Engineering and Technology, June Issue