Behavior recognition is already starting to move towards some scenarios, although it needs to be more sophisticated technically.

       According to the "2019-2024 Research Report on China's Machine Vision Industry Prospects and Investment Opportunities" released by the China Business Industry Research Institute, the scale of China's machine vision market exceeded 10 billion yuan for the first time in 2018; , the machine vision market will further expand,It is estimated that the machine vision market will reach nearly 12.5 billion yuan in 2019.

      It is true that the market size of the CV (machine vision) industry is not small and profitable, but when technology products mature and start to be applied, how to eat this cake has become the biggest problem faced by many CV startups. . At the same time, continuous losses and profit pressure are also urging every CV company to “run”.

       趨視科技并不屬于CV領(lǐng)域最知名的行業(yè)，然而它們卻在落地應(yīng)用和盈利上先人一步，其公司創(chuàng)始人徐飆表示："If 90% of companies in the industry are losing money, we belong to the other 10%."

        how do they do it

Figure | Xu Biao, founder of Truthvision Technology

CV not only has face recognition, but also behavior analysis

      When it comes to CV, attention and topics are often concentrated in the field of face recognition. SenseTime, Megvii, etc. are the focus of attention both inside and outside the industry, but CV is not equivalent to face recognition, it also includes behavior recognition. Xu Biao introduced that since its establishment, Truthvision Technology has always aimed at behavior recognition.

       “行為識(shí)別就是識(shí)別人類或者車的行為，Such as people's fighting behavior, car running red light behavior and so on. Although both belong to machine vision, face recognition and behavior recognition are two technologies and different fields. "

At the technical level, face recognition can be completed through a photo, while behavior recognition needs to be judged by combining continuous data, because behavior itself is a continuous and dynamic process.in short,Face recognition solves the problem of who the target is, and behavior solves what kind of thing.At present, behavior recognition is often used in judicial management, smart stores,intelligent

       Xu Biao told us: "There are many fields where behavior recognition is applicable, but because the technology is not mature enough, it is difficult for behavior recognition technology to play a very good role in the face of too complex and non-standard scenes. Therefore, this technology can only It is first applied in some vertical scenarios, and gradually accumulated and improved in the process of application, so as to expand to more scenarios, and finally meet the requirements of human behavior cognition in a large range."

        So what are the technical difficulties of behavior recognition?

        Since behavior is diverse, it includes individual behavior and group behavior, and each behavior is expressed in different ways. For example, fighting and stealing, fighting between individuals and fighting between groups are completely different.Therefore, behavior recognition faces great difficulties at the data collection level, which mainly involve problems such as occlusion and dislocation.

        At the same time, the angle of human viewing the world is three-dimensional, and the picture captured by the camera is two-dimensional, so there will be a person in the video showing an arm, but because the distance parameter cannot be collected in the video,所以遮擋、錯(cuò)位的現(xiàn)象會(huì)讓AI算法難以判斷。

       其次學(xué)習(xí)數(shù)據(jù)欠缺。眾所周知，許多AI技術(shù)依靠深度學(xué)習(xí)算法模型去訓(xùn)練，這導(dǎo)致要讓AI實(shí)現(xiàn)行為識(shí)別，就必須先給行為下定義，讓AI知道行為是什么。然而前面已經(jīng)提到行為非常復(fù)雜，甚至很多時(shí)候AI需要學(xué)習(xí)判斷的是負(fù)面行為，因此企業(yè)很難獲取到大量的學(xué)習(xí)數(shù)據(jù)。而算法模型沒有經(jīng)過大量數(shù)據(jù)去訓(xùn)練，也就很難“聰明”起來，從而在識(shí)別的效果和精度上難以達(dá)到用戶需求。

        不過盡管在技術(shù)上需要更加精進(jìn)，但行為識(shí)別已經(jīng)開始走向一些場景。

CV企業(yè)破冰關(guān)鍵：規(guī)模化

       徐飆介紹：“公司一開始關(guān)注的就是行業(yè)落地而非通用場景，且瞄準(zhǔn)的第一個(gè)領(lǐng)域就是司法領(lǐng)域行業(yè)的管理，比如監(jiān)獄管理，是否有犯人打斗、翻墻、攀爬等。這對(duì)于司法領(lǐng)域的管理而言是一個(gè)剛需，能夠降低人力管理成本，提升管理質(zhì)量。”

       而行業(yè)落地和通用場景落地兩條路徑的最大區(qū)別，在徐飆看來，前者能夠助力企業(yè)快速實(shí)現(xiàn)規(guī)模化落地，而這至關(guān)重要。

       他談到：“所有CV廠商在近年來特別強(qiáng)調(diào)落地，本質(zhì)上就是規(guī)模化落地，即企業(yè)在一個(gè)項(xiàng)目試點(diǎn)實(shí)現(xiàn)技術(shù)落地后能夠快速復(fù)制到下一個(gè)同類型的場景中，而不是做完一個(gè)試點(diǎn)，下一個(gè)場景再重新做一遍，這無疑增加了許多成本。”

       對(duì)于企業(yè)而言，要實(shí)現(xiàn)規(guī)模化落地首先在最初尋找落地行業(yè)時(shí)，就要找到能夠?qū)崿F(xiàn)規(guī)模化、可復(fù)制性強(qiáng)的場景。其中的關(guān)鍵在于，企業(yè)對(duì)于用戶核心訴求的把握是否精確。徐飆認(rèn)為，CV企業(yè)要實(shí)現(xiàn)規(guī)模化必須了解用戶的需求，所謂需求指的不僅是用戶對(duì)于功能的需求，還包括用戶對(duì)性能當(dāng)中準(zhǔn)確度的要求。

       “這需要碰撞。有些時(shí)候沒有人會(huì)告訴你他的需求和對(duì)準(zhǔn)確度的要求是什么，企業(yè)往往需要通過試點(diǎn)、交流、反饋、修正......逐步形成一個(gè)行業(yè)共識(shí)，而并非單個(gè)客戶的需求。”

       但即便把握了用戶需求和性能指標(biāo)并不足夠，企業(yè)還要評(píng)估自身的技術(shù)體系、優(yōu)勢能否滿足用戶的需求和指標(biāo)。最后企業(yè)還要考慮實(shí)現(xiàn)規(guī)模化之后，是否會(huì)被競品取代，這要求其必須在技術(shù)落地應(yīng)用過程中打造自身的技術(shù)門檻，如此廠商們才能率先占領(lǐng)市場，并在后續(xù)的競爭中獲勝。

       回到趨視科技自身，徐飆談道：“公司明年的短期計(jì)劃，一方面是確保在司法行業(yè)實(shí)現(xiàn)規(guī)模化，創(chuàng)造更多的收益；同時(shí)也會(huì)將技術(shù)落地到智慧門店場景。小規(guī)模化帶給我們盈利，也驗(yàn)證了技術(shù)已經(jīng)達(dá)到可復(fù)制狀態(tài)，所以我們將會(huì)向更大的市場進(jìn)行布局。”

【鎂客·請(qǐng)講】欄目   策劃&撰寫：溫暖