Class Activation Mapping-Based Car Saliency Region and Detection for In-Vehicle Surveillance

This paper describes a class of problem where a detection and localization of car are addressed simultaneously using in-vehicle camera. While the current state-of-the-art researches have tackled the detection problems, in fact, there several rooms for the improvement. One of them is to optimize the detection processes. In this research, we introduce the use of activation map obtained from the last layer of the convolutional network to localize the saliency region of the car and to perform detection of the car objects by collaborating deep architecture trained weight. Our study also finds the benefit of changing the last layer of the network, from fully-connected layer to Global Average Pooling (GAP), which alleviates the overfitting problem. Experiment results on various optimization methods show the impact of the performance, where the proposed GAP-based deep learning and ADAM optimization gives the best results, up to 99.83%. It imply the possibilities of utilizing our method for real case usage scenario.