Frame Similarity Assessment on the Edge for Improved Object Detection

Real-time object detection is a recurring challenge in the field of computer vision. It is an important task for many real-time applications, such as trespassing detection or autonomous driving. However, real-time processing requires high computational power, increasing the latency of object detection. Delays can be a critical problem for constrained real-time applications running on the edge, such as advanced driving assistance. Therefore, this work proposes a method, called Lightweight Pixel Difference Accumulation (LWPDA), to reduce the processing effort by frames which are similar to recently processed ones. To do this, frames are compared in order to estimate their similarity. In this work, we compare our method against five different one: Mean Squared Error (MSE), Structural Similarity Index Measure (SSIM), ImageHash and Histogram comparison. Our results show that two of the methods stand out: ImageHash and LWPDA. The first one was able to reduce the processing time per frame by approximately 35%, at the cost of 13% accuracy, while LWPDA reduced processing by 40%, at the same cost of 13% accuracy.