Context: Object detection is gaining more influence in everyday life, various institutions and agencies are utilizing it to help streamline their day-to-day tasks. It helps process large quantities of data and requires less resources, hence making for a promising tool in the future. It still faces many baseline issues, such as weather conditions obstructing the shape or tone of an object and thereby causing misidentifications. This can be as harmless as a toy misidentifying a frown for a smile and reacting happily to your sorrow, or as harmful as a self-driving car misidentifying a dark car. If for instance the darker car’s lights were in disrepair, and it was placed under the shade of a few trees at a crossing, the first car might continue through when it should have slowed down in good time or even stopped to ensure it can prevent an accident. The intent of this research is to delve into those aspects and scenarios where the weather and natural lighting outdoors can affect object detection in traffic from the perspective of a police vehicle’s camera. Evidence of law enforcement attempting implementation of the technology can be readily found on the internet even as far back as 2010, providing the right relevance for this study.

Realization (Method): The research will be conducted using four common categories of objects encountered in everyday traffic; cars, people, motorbikes and trains. Each category will have three instances in form of images in their relevant setting, on streets for instance, to represent them in the conducted tests. Each test consists of four filters; contrast, blur, noise and resizing. For each filter there will be 20 versions, i.e. every fifth degree will be an option to apply and these will all be combined to make a total of 20^4 combinations for each image, then all combinations will be tested and detections will be registered.

Objectives: The scope of objectives for this study was to find out which of the four categories was easiest to detect, which of the four filters was most disruptive, and to find out if there are any rules of thumb for what degrees of each filter could be considered a threshold beyond which detection is not guaranteed.

Results: The results proved cars and people to be easiest to detect, noise to be the most obstructive filter, and contrast to guarantee detection up to 30% of application from the original luminance. Blur and change of size were negligible in impact and thus did not matter, while noise was too complex to give a clear answer in regards to beyond what percentage of noise application stops all further detections.

Conclusions: What could be concluded from this study is that certain visual effects are harmless, contrast and noise are predominant, and that more research into the disruption of noise should be done. Noise meaning particles or specks of black or white color in the shape of pixels strewn across an image(i.e. “Salt-and-pepper noise”) to simulate things such as snow, rain etc. Object detection has been costly in mistakes even very recently in the public sphere so it needs more optimization. But it has many uses in many fields such as medicine, law enforcement, statistics and the fire departments, and for broader, commercial use, models need more training.