Efficient and Fast Objects Detection Technique for Intelligent Video Surveillance Using Transfer Learning and Fine-Tuning

Abstract

Intelligent video surveillance systems require effective techniques in order to detect objects accurately and rapidly. The most suitable algorithms for performing this task are based on convolutional neural networks. Existing approaches encounter a wide range of difficulties in terms of dealing with different sizes, high definition, or colored images turning these latter slower and less precise. The real-time sensitive application offers an interesting challenge for the optimization of the quality and quantity of previous approaches, thus obtaining an efficient system with regard to surveillance environment. This paper presents a novel, fast, and precise technique for advanced object detection as far as intelligent video surveillance systems are concerned. Thus, we propose the transfer learning of an efficient pre-trained network to appropriate datasets for our application and its integration in the architecture of our algorithm. Accordingly, we implement a fine-tuning on this pre-trained model via replacing the softmax layer and running backpropagation. Then, we compare the results of the previous algorithms using common evaluation parameters. The experimental results reveal that with this technique, we can enhance the precision and the accuracy of object detection in video surveillance scenes to more than $$90 \%$$. Furthermore, along with dealing with different input dimensions, the detector runs in real time. To conclude, our application of machine learning for intelligent video surveillance systems maximizes their efficiency in highly difficult situations.