We are developing a PCB (printed circuit board) inspection system that can detect correct orientation and type of IC chips using webcam and shape detection algorithms. It is also integrated with a projection system to indicate the correct position of the IC in a PCB. Our system can be used without any CAD description of the IC chip and based on classical computer vison based image processing algorithm that does not require prior training. In particular, we evaluated our algorithm to detect 19 different integrated circuits (IC) from 10 different printed circuit boards (PCB) of different colors. We have compared three different shape descriptors for four different color space models. We have evaluated shape detection algorithms in different lighting conditions (indoor, outdoor, and controlled light source) to find suitable illumination for image acquisition. We undertook statistical hypothesis testing to find the effect of color space models and shape descriptors on the accuracy, false positive and false negative rates. While measuring accuracy, we have noted that L*a*b* color space is significantly worse, and the best result is obtained in YCbCr color space using bounding box shape descriptors for 2500 Lux using LED. [Paper link: 10.1017/S0890060419000398]
Lane detection is a visualization-based driver assistance system for lane departure warning, parking assistance, off-road warning, and lane keeping assistance. There is a plethora of lane detection models and lane datasets exist. However, lane detection in unstructured environment is still a challenging problem. We address both these problems by creating an unstructured and challenging lane dataset and developing lane detection model. With a significant improvement of 42% over existing state-of-the-art lane detection model, this technology can be used for assisting autonomous driving system. It also can be used for other problems like for autonomous taxiing of aircraft as we have shown in supplementary video. [Paper Link: 10.1109/TAI.2022.3212347]
The ongoing Covid-19 pandemic has made it challenging for large scale data collection, in particular for Convolutional Neural Network (CNN)-based computer vision systems. Additionally, there are numerous circumstances where security, privacy, and limitations pertaining to the accessibility of the required equipment make it arduous to validate computer vision systems with real-world datasets. In this work, we investigated the possibilities of using synthetic datasets, generated from Virtual Environments (VE) for training and validation of CNN models. We present two use cases where the above-mentioned circumstances play a vital role in preparing the datasets and validating the model with large-scale datasets. By developing and leveraging a three-dimensional Digital Twin (DT), we produce large scale datasets for validating social distancing in workspaces; and in the context of semi-autonomous vehicles, we evaluate how a CNN-based object detection model would perform in an Indian road scenario.
[Paper link: 10.1016/j.vrih.2022.01.004]
This work focusing on developing object detection model and comparing it with state-of-the-art models. It used lane detection model and formulated as multi-class semantic segmentation problem with few modification in the architecture. To train on unstructured road environments, Indian Driving Dataset (IDD) was chosen with selected seven class (autorickshaw, bus, car, motorcycle, person, rider, truck) of objects. The proposed model and nine other state-of-the art models (including two-stage, one-stage, transformer, and segmentation based) were evaluated on IDD and was found that the proposed model achieved better accuracy (mean IoU of 0.36) than other models. Finally, a detailed ablation study was conducted to understand the superiority of the proposed architecture than its individual branches. It was also tested with different popular loss functions (mean square error, weighted cross entropy, categorical cross entropy, and shape aware loss function) and was found that categorical cross entropy showed better performance than others. [Paper link: 10.1007/s11760-024-03590-7]