Object Detection from Video Using Temporal Convolutional Network
ELizabeth Sabu, K. Suresh · 2018
The exciting possibilities for tackling many video understanding problems opened up by the increasing availability of video sensors and high performance video processing hardware. This reveals the need of a robust and flexible video object detection technique. In this work, various techniques for object detection and recognition are studied and a novel technique is developed. Method using R-CNN, AlexNet and MobileNet are some of the main techniques. Since object detection on videos dealing with large temporal fluctuations is challenging. The model used for this work is created with Inception_V2as base architecture with Faster R-CNN convolutional layers and with TCN layers which is a better video detection algorithm. The temporal convolution network is proposed to incorporate temporal information to regularize the detection results and shows its effectiveness for the task. The performance is compared with the existing two methods and it shows better results than still image object detection methods. The improved method introduced in this work can thereby be extremely useful to detect the objects even in the presence of inter-frame temporal fluctuations.