Computer Vision Algorithms For Detection Research Articles

Seeing is caring – automated assessment of resource use of broilers with computer vision techniques

Routine monitoring of broiler chickens provides insights in the welfare status of a flock, helps to guarantee minimum defined levels of animal welfare and assists farmers in taking remedial measures at an early stage. Computer vision techniques offer exciting potential for routine and automated assessment of broiler welfare, providing an objective and biosecure alternative to the current more subjective and time-consuming methods. However, the current state-of-the-art computer vision solutions for assessing broiler welfare are not sufficient to allow the transition to fully automated monitoring in a commercial environment. Therefore, the aim of this study was to investigate the potential of computer vision algorithms for detection and resource use monitoring of broilers housed in both experimental and commercial settings, while also assessing the potential for scalability and resource-efficient implementation of such solutions. This study used a combination of detection and resource use monitoring methods, where broilers were first detected using Mask R-CNN and were then assigned to a specific resource zone using zone-based classifiers. Three detection models were proposed using different annotation datasets: model A with annotated broilers from a research facility, model B with annotated broilers from a commercial farm, and model A+B where annotations from both environments were combined. The algorithms developed for individual broiler detection performed well for both the research facility (model A, F1 score > 0.99) and commercial farm (model A+B, F1 score > 0.83) test data with an intersection over union of 0.75. The subsequent monitoring of resource use at the commercial farm using model A+B for broiler detection, also performed very well for the feeders, bale and perch (F1 score > 0.93), but not for the drinkers (F1 score = 0.28), which was likely caused by our evaluation method. Thus, the algorithms used in this study are a first step to measure resource use automatically in commercial application and allow detection of a large number of individual animals in a non-invasive manner. From location data of every frame, resource use can be calculated. Ultimately, the broiler detection and resource use monitoring might further be used to assess broiler welfare.

Deep Learning-Based Water Segmentation for Autonomous Surface Vessel

Visual-based obstacle detection from an autonomous surface vessel (ASV) is a complex task due to high variance of scene properties such as different illumination and presence of reflections. One approach in implementing the task is through extracting waterlines to enable inferring of vessel orientation and obstacles presence. Classical computer vision algorithms for detection holds limitation in robustness and scalability. With recent breakthroughs in deep neural network architectures, vision-based object detection is seen to obtain high performance. In this work, the deep learning models based on Convolutional Neural Network (CNN) to implement binary semantic segmentation is studied. This architecture identifies each pixel to water and non-water classes. In purpose of benchmarking models, Fully Convolutional Network (FCN), SegNet and U-Net are trained on a publicly available dataset, IntCatch Vision Data Set (ICVDS), to evaluate the performance. From the experiments carried out, quantitative results show effectiveness of the models with accuracy all above 95.55% and lowest average speed of 11 frames per second. To improve, pre-trained networks (VGG 16, Resnet-50 and MobileNet) are used as a backbone, obtaining an improved accuracy above 98.14% with lowest inferring speed of 10 frame per second. Using the developed ASV, new dataset of 143 images called Malaysia ASV Dataset (MASVD) is collected, labelled and made publicly available. The trained models are tested with the newly collected dataset obtaining accuracy of 75%. The high accuracy performance shows potential for the models to be employed for collision avoidance algorithm in ASV navigation.

Evaluation of Multiclass Traffic Sign Detection and Classification Methods for U.S. Roadway Asset Inventory Management

Frequent analysis and updating the condition of traffic signs and mile markers are among the most important aspects of a highway asset management system. Today’s practices mainly involve manual data collection and analysis, which have to be done for millions of miles of roads and the practice needs to be repeated regularly. While significant progress has been made on improving the data collection practice by leveraging video streams collected from car-mounted cameras, the analysis has primarily remained a manual and labor-intensive process. Automating the analysis from the collected videos is also challenging due to the interclass variability of traffic signs, expected changes in illumination, occlusion, sign position, and orientation. To address these challenges, this paper presents and evaluates the performance of three computer vision algorithms for detection and classification of traffic signs in presence of cluttered backgrounds and static and dynamic occlusions. The task particularly focuses on (1) extracting two-dimensional (2D) candidate windows from already collected video streams that potentially contain traffic signs–without making any prior assumption about their locations; (2) detecting the presence of signs in these 2D candidate windows; and (3) classifying them into warning, regulatory, stop, and yield sign categories based on their shape and color. For validation, a new comprehensive benchmark data set of over 11,000 annotated U.S. traffic sign images with a large range of pose, scale, background, illumination, and occlusion variation is introduced. Experimental results show an average accuracy of 76.20%, 89.31%, and 94.83% for the methods of (1) Haar-like features with Cascade classifiers, (2) histograms of oriented gradients (HOG) with multiple one-versus-all support vector machine (SVM) classifiers, and (3) HOG+C with the SVM classifiers, a variant of the second method with histograms of colors concatenated to HOG. The experimental results demonstrate the potential of leveraging joint representation of texture and color in HOG+C together with SVM discriminative classifiers as a viable solution for creating up-to-date and complete inventories of traffic signs for U.S. roadways.

Multi Object Detection and Tracking from Video File

This paper describes computer vision algorithms for detection, identification, and tracking of moving objects in a video file. The problem of multiple object tracking can be divided into two parts; detecting moving objects in each frame and associating the detections corresponding to the same object over time. The detection of moving objects uses a background subtraction algorithm based on Gaussian mixture models. The motion of each track is estimated by a Kalman filter. The video tracking algorithm was successfully tested using the BIWI walking pedestrians datasets [. The experimental results show that system can operate in real time and successfully detect, track and identify multiple targets in the presence of partial occlusion.

Smart video surveillance for airborne platforms

This paper describes real-time computer vision algorithms for detection, identification, and tracking of moving targets in video streams generated by a moving airborne platform. Moving platforms cause instabilities in image acquisition due to factors such as disturbances and the ego-motion of the camera that distorts the actual motion of the moving targets. When the camera is mounted on a moving observer, the entire scene (background and targets) appears to be moving and the actual motion of the targets must be separated from the background motion. The motion of the airborne platform is modeled as affine transformation and its parameters are estimated using corresponding feature sets in consecutive images. After motion is compensated, the platform is considered as stationary and moving targets are detected accordingly. A number of tracking algorithms including particle filters, mean-shift, and connected component were implemented and compared. A cascaded boosted classifier with Haar wavelet feature extraction for moving target classification was developed and integrated with the recognition system that uses joint-feature spatial distribution. The integrated smart video surveillance system has been successfully tested using the Vivid Datasets provided by the Air Force Research Laboratory. The experimental results show that system can operate in real time and successfully detect, track, and identify multiple targets in the presence of partial occlusion.