QAOVDetect: A Novel Syllogistic Model with Quantized and Anchor Optimized Approach to Assist Visually Impaired for Animal Detection using 3D Vision

Abstract

In developing countries, stray animals can be frequently encountered on the roads, pathways, campuses, and other places. Due to this, the visually impaired (VI) are at more risk than the sighted ones. To gain more security and safety, they need a solution to deal with their problem. This experimentation aims to develop a hardware–software integrated solution that can detect stray animals. Along with the detection, the solution should also provide the distance of those objects from the user and alert them if it is getting closer. To put this experiment together, Jetson NANO and ZED mini camera have been chosen for processing and image capturing to make the solution mobile and accurate as they are leading hardware devices. A novel approach involving quantization and anchor-optimization has been proposed using Single-Shot Detector (SSD) Resnet 50 FPN as the base model. The model has been compressed by quantization to reduce the inference time, and anchor optimization has been done to compensate for the accuracy loss faced during quantization. We have performed experimentation by training the original model, anchor-optimized model, and quantization plus anchor-optimized model using batch sizes 64 and 8. This experimentation has been done to understand the effect of anchor-optimization and quantization on the base model and the effect of the batch size used for training on different model versions. The performance of all the models with applied quantization and anchor-optimization for both batch sizes 64 and 8 has been noted in mAP. The mAP of the quantized plus anchor-optimized model trained using batch size 64 was the highest, i.e. 93.5%. It can also be concluded that we can achieve the light-weight model with the best performance by balancing quantization and anchor-optimization to make it suitable for an edge device using batch size 64. © 2022, The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature.